Linux Audio

Check our new training course

Loading...
v3.1
 
   1/*
   2 *  linux/mm/page_alloc.c
   3 *
   4 *  Manages the free list, the system allocates free pages here.
   5 *  Note that kmalloc() lives in slab.c
   6 *
   7 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   8 *  Swap reorganised 29.12.95, Stephen Tweedie
   9 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
  10 *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
  11 *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
  12 *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  13 *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  14 *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
  15 */
  16
  17#include <linux/stddef.h>
  18#include <linux/mm.h>
  19#include <linux/swap.h>
  20#include <linux/interrupt.h>
  21#include <linux/pagemap.h>
  22#include <linux/jiffies.h>
  23#include <linux/bootmem.h>
  24#include <linux/memblock.h>
  25#include <linux/compiler.h>
  26#include <linux/kernel.h>
  27#include <linux/kmemcheck.h>
 
  28#include <linux/module.h>
  29#include <linux/suspend.h>
  30#include <linux/pagevec.h>
  31#include <linux/blkdev.h>
  32#include <linux/slab.h>
  33#include <linux/ratelimit.h>
  34#include <linux/oom.h>
  35#include <linux/notifier.h>
  36#include <linux/topology.h>
  37#include <linux/sysctl.h>
  38#include <linux/cpu.h>
  39#include <linux/cpuset.h>
  40#include <linux/memory_hotplug.h>
  41#include <linux/nodemask.h>
  42#include <linux/vmalloc.h>
  43#include <linux/vmstat.h>
  44#include <linux/mempolicy.h>
  45#include <linux/stop_machine.h>
  46#include <linux/sort.h>
  47#include <linux/pfn.h>
  48#include <linux/backing-dev.h>
  49#include <linux/fault-inject.h>
  50#include <linux/page-isolation.h>
  51#include <linux/page_cgroup.h>
  52#include <linux/debugobjects.h>
  53#include <linux/kmemleak.h>
  54#include <linux/memory.h>
  55#include <linux/compaction.h>
  56#include <trace/events/kmem.h>
  57#include <linux/ftrace_event.h>
  58#include <linux/memcontrol.h>
  59#include <linux/prefetch.h>
  60
  61#include <asm/tlbflush.h>
 
 
 
 
 
 
 
 
 
 
 
  62#include <asm/div64.h>
  63#include "internal.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  64
  65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
  66DEFINE_PER_CPU(int, numa_node);
  67EXPORT_PER_CPU_SYMBOL(numa_node);
  68#endif
  69
 
 
  70#ifdef CONFIG_HAVE_MEMORYLESS_NODES
  71/*
  72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
  73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
  74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
  75 * defined in <linux/topology.h>.
  76 */
  77DEFINE_PER_CPU(int, _numa_mem_);		/* Kernel "local memory" node */
  78EXPORT_PER_CPU_SYMBOL(_numa_mem_);
  79#endif
  80
 
 
 
 
 
 
 
  81/*
  82 * Array of node states.
  83 */
  84nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
  85	[N_POSSIBLE] = NODE_MASK_ALL,
  86	[N_ONLINE] = { { [0] = 1UL } },
  87#ifndef CONFIG_NUMA
  88	[N_NORMAL_MEMORY] = { { [0] = 1UL } },
  89#ifdef CONFIG_HIGHMEM
  90	[N_HIGH_MEMORY] = { { [0] = 1UL } },
  91#endif
 
  92	[N_CPU] = { { [0] = 1UL } },
  93#endif	/* NUMA */
  94};
  95EXPORT_SYMBOL(node_states);
  96
  97unsigned long totalram_pages __read_mostly;
  98unsigned long totalreserve_pages __read_mostly;
  99int percpu_pagelist_fraction;
 100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
 101
 102#ifdef CONFIG_PM_SLEEP
 103/*
 104 * The following functions are used by the suspend/hibernate code to temporarily
 105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
 106 * while devices are suspended.  To avoid races with the suspend/hibernate code,
 107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
 108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
 109 * guaranteed not to run in parallel with that modification).
 110 */
 111
 112static gfp_t saved_gfp_mask;
 113
 114void pm_restore_gfp_mask(void)
 115{
 116	WARN_ON(!mutex_is_locked(&pm_mutex));
 117	if (saved_gfp_mask) {
 118		gfp_allowed_mask = saved_gfp_mask;
 119		saved_gfp_mask = 0;
 120	}
 121}
 122
 123void pm_restrict_gfp_mask(void)
 124{
 125	WARN_ON(!mutex_is_locked(&pm_mutex));
 126	WARN_ON(saved_gfp_mask);
 127	saved_gfp_mask = gfp_allowed_mask;
 128	gfp_allowed_mask &= ~GFP_IOFS;
 129}
 130#endif /* CONFIG_PM_SLEEP */
 131
 132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 133int pageblock_order __read_mostly;
 134#endif
 135
 136static void __free_pages_ok(struct page *page, unsigned int order);
 
 137
 138/*
 139 * results with 256, 32 in the lowmem_reserve sysctl:
 140 *	1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 141 *	1G machine -> (16M dma, 784M normal, 224M high)
 142 *	NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 143 *	HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 144 *	HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
 145 *
 146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 147 * don't need any ZONE_NORMAL reservation
 148 */
 149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
 150#ifdef CONFIG_ZONE_DMA
 151	 256,
 152#endif
 153#ifdef CONFIG_ZONE_DMA32
 154	 256,
 155#endif
 
 156#ifdef CONFIG_HIGHMEM
 157	 32,
 158#endif
 159	 32,
 160};
 161
 162EXPORT_SYMBOL(totalram_pages);
 163
 164static char * const zone_names[MAX_NR_ZONES] = {
 165#ifdef CONFIG_ZONE_DMA
 166	 "DMA",
 167#endif
 168#ifdef CONFIG_ZONE_DMA32
 169	 "DMA32",
 170#endif
 171	 "Normal",
 172#ifdef CONFIG_HIGHMEM
 173	 "HighMem",
 174#endif
 175	 "Movable",
 
 
 
 176};
 177
 178int min_free_kbytes = 1024;
 
 
 
 
 
 
 
 
 
 
 
 179
 180static unsigned long __meminitdata nr_kernel_pages;
 181static unsigned long __meminitdata nr_all_pages;
 182static unsigned long __meminitdata dma_reserve;
 183
 184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
 185  /*
 186   * MAX_ACTIVE_REGIONS determines the maximum number of distinct
 187   * ranges of memory (RAM) that may be registered with add_active_range().
 188   * Ranges passed to add_active_range() will be merged if possible
 189   * so the number of times add_active_range() can be called is
 190   * related to the number of nodes and the number of holes
 191   */
 192  #ifdef CONFIG_MAX_ACTIVE_REGIONS
 193    /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
 194    #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
 195  #else
 196    #if MAX_NUMNODES >= 32
 197      /* If there can be many nodes, allow up to 50 holes per node */
 198      #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
 199    #else
 200      /* By default, allow up to 256 distinct regions */
 201      #define MAX_ACTIVE_REGIONS 256
 202    #endif
 203  #endif
 204
 205  static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
 206  static int __meminitdata nr_nodemap_entries;
 207  static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
 208  static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
 209  static unsigned long __initdata required_kernelcore;
 210  static unsigned long __initdata required_movablecore;
 211  static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
 212
 213  /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 214  int movable_zone;
 215  EXPORT_SYMBOL(movable_zone);
 216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
 217
 218#if MAX_NUMNODES > 1
 219int nr_node_ids __read_mostly = MAX_NUMNODES;
 220int nr_online_nodes __read_mostly = 1;
 221EXPORT_SYMBOL(nr_node_ids);
 222EXPORT_SYMBOL(nr_online_nodes);
 223#endif
 224
 
 
 
 
 
 
 225int page_group_by_mobility_disabled __read_mostly;
 226
 227static void set_pageblock_migratetype(struct page *page, int migratetype)
 
 
 
 
 
 
 
 
 228{
 
 
 229
 230	if (unlikely(page_group_by_mobility_disabled))
 231		migratetype = MIGRATE_UNMOVABLE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 232
 233	set_pageblock_flags_group(page, (unsigned long)migratetype,
 234					PB_migrate, PB_migrate_end);
 
 235}
 236
 237bool oom_killer_disabled __read_mostly;
 
 
 
 
 
 
 
 
 238
 239#ifdef CONFIG_DEBUG_VM
 240static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 241{
 242	int ret = 0;
 243	unsigned seq;
 244	unsigned long pfn = page_to_pfn(page);
 
 245
 246	do {
 247		seq = zone_span_seqbegin(zone);
 248		if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
 249			ret = 1;
 250		else if (pfn < zone->zone_start_pfn)
 251			ret = 1;
 252	} while (zone_span_seqretry(zone, seq));
 253
 
 
 
 
 
 254	return ret;
 255}
 256
 257static int page_is_consistent(struct zone *zone, struct page *page)
 258{
 259	if (!pfn_valid_within(page_to_pfn(page)))
 260		return 0;
 261	if (zone != page_zone(page))
 262		return 0;
 263
 264	return 1;
 265}
 266/*
 267 * Temporary debugging check for pages not lying within a given zone.
 268 */
 269static int bad_range(struct zone *zone, struct page *page)
 270{
 271	if (page_outside_zone_boundaries(zone, page))
 272		return 1;
 273	if (!page_is_consistent(zone, page))
 274		return 1;
 275
 276	return 0;
 277}
 278#else
 279static inline int bad_range(struct zone *zone, struct page *page)
 280{
 281	return 0;
 282}
 283#endif
 284
 285static void bad_page(struct page *page)
 286{
 287	static unsigned long resume;
 288	static unsigned long nr_shown;
 289	static unsigned long nr_unshown;
 290
 291	/* Don't complain about poisoned pages */
 292	if (PageHWPoison(page)) {
 293		reset_page_mapcount(page); /* remove PageBuddy */
 294		return;
 295	}
 296
 297	/*
 298	 * Allow a burst of 60 reports, then keep quiet for that minute;
 299	 * or allow a steady drip of one report per second.
 300	 */
 301	if (nr_shown == 60) {
 302		if (time_before(jiffies, resume)) {
 303			nr_unshown++;
 304			goto out;
 305		}
 306		if (nr_unshown) {
 307			printk(KERN_ALERT
 308			      "BUG: Bad page state: %lu messages suppressed\n",
 309				nr_unshown);
 310			nr_unshown = 0;
 311		}
 312		nr_shown = 0;
 313	}
 314	if (nr_shown++ == 0)
 315		resume = jiffies + 60 * HZ;
 316
 317	printk(KERN_ALERT "BUG: Bad page state in process %s  pfn:%05lx\n",
 318		current->comm, page_to_pfn(page));
 319	dump_page(page);
 320
 
 321	dump_stack();
 322out:
 323	/* Leave bad fields for debug, except PageBuddy could make trouble */
 324	reset_page_mapcount(page); /* remove PageBuddy */
 325	add_taint(TAINT_BAD_PAGE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 326}
 327
 328/*
 329 * Higher-order pages are called "compound pages".  They are structured thusly:
 330 *
 331 * The first PAGE_SIZE page is called the "head page".
 332 *
 333 * The remaining PAGE_SIZE pages are called "tail pages".
 
 334 *
 335 * All pages have PG_compound set.  All pages have their ->private pointing at
 336 * the head page (even the head page has this).
 337 *
 338 * The first tail page's ->lru.next holds the address of the compound page's
 339 * put_page() function.  Its ->lru.prev holds the order of allocation.
 340 * This usage means that zero-order pages may not be compound.
 341 */
 342
 343static void free_compound_page(struct page *page)
 344{
 345	__free_pages_ok(page, compound_order(page));
 346}
 347
 348void prep_compound_page(struct page *page, unsigned long order)
 349{
 350	int i;
 351	int nr_pages = 1 << order;
 352
 353	set_compound_page_dtor(page, free_compound_page);
 354	set_compound_order(page, order);
 355	__SetPageHead(page);
 356	for (i = 1; i < nr_pages; i++) {
 357		struct page *p = page + i;
 358
 359		__SetPageTail(p);
 360		p->first_page = page;
 361	}
 362}
 363
 364/* update __split_huge_page_refcount if you change this function */
 365static int destroy_compound_page(struct page *page, unsigned long order)
 366{
 367	int i;
 368	int nr_pages = 1 << order;
 369	int bad = 0;
 370
 371	if (unlikely(compound_order(page) != order) ||
 372	    unlikely(!PageHead(page))) {
 373		bad_page(page);
 374		bad++;
 375	}
 376
 377	__ClearPageHead(page);
 
 378
 379	for (i = 1; i < nr_pages; i++) {
 380		struct page *p = page + i;
 
 381
 382		if (unlikely(!PageTail(p) || (p->first_page != page))) {
 383			bad_page(page);
 384			bad++;
 385		}
 386		__ClearPageTail(p);
 387	}
 
 
 
 
 388
 389	return bad;
 
 
 
 390}
 391
 392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
 
 
 393{
 394	int i;
 
 
 
 
 
 
 395
 396	/*
 397	 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
 398	 * and __GFP_HIGHMEM from hard or soft interrupt context.
 
 
 399	 */
 400	VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
 401	for (i = 0; i < (1 << order); i++)
 402		clear_highpage(page + i);
 
 
 403}
 404
 405static inline void set_page_order(struct page *page, int order)
 
 406{
 407	set_page_private(page, order);
 408	__SetPageBuddy(page);
 409}
 410
 411static inline void rmv_page_order(struct page *page)
 
 
 412{
 413	__ClearPageBuddy(page);
 414	set_page_private(page, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 415}
 416
 417/*
 418 * Locate the struct page for both the matching buddy in our
 419 * pair (buddy1) and the combined O(n+1) page they form (page).
 420 *
 421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
 422 * the following equation:
 423 *     B2 = B1 ^ (1 << O)
 424 * For example, if the starting buddy (buddy2) is #8 its order
 425 * 1 buddy is #10:
 426 *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 427 *
 428 * 2) Any buddy B will have an order O+1 parent P which
 429 * satisfies the following equation:
 430 *     P = B & ~(1 << O)
 431 *
 432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
 433 */
 434static inline unsigned long
 435__find_buddy_index(unsigned long page_idx, unsigned int order)
 436{
 437	return page_idx ^ (1 << order);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 438}
 439
 440/*
 441 * This function checks whether a page is free && is the buddy
 442 * we can do coalesce a page and its buddy if
 443 * (a) the buddy is not in a hole &&
 444 * (b) the buddy is in the buddy system &&
 445 * (c) a page and its buddy have the same order &&
 446 * (d) a page and its buddy are in the same zone.
 447 *
 448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
 449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
 450 *
 451 * For recording page's order, we use page_private(page).
 452 */
 453static inline int page_is_buddy(struct page *page, struct page *buddy,
 454								int order)
 
 455{
 456	if (!pfn_valid_within(page_to_pfn(buddy)))
 457		return 0;
 458
 459	if (page_zone_id(page) != page_zone_id(buddy))
 460		return 0;
 461
 462	if (PageBuddy(buddy) && page_order(buddy) == order) {
 463		VM_BUG_ON(page_count(buddy) != 0);
 464		return 1;
 465	}
 466	return 0;
 467}
 468
 469/*
 470 * Freeing function for a buddy system allocator.
 471 *
 472 * The concept of a buddy system is to maintain direct-mapped table
 473 * (containing bit values) for memory blocks of various "orders".
 474 * The bottom level table contains the map for the smallest allocatable
 475 * units of memory (here, pages), and each level above it describes
 476 * pairs of units from the levels below, hence, "buddies".
 477 * At a high level, all that happens here is marking the table entry
 478 * at the bottom level available, and propagating the changes upward
 479 * as necessary, plus some accounting needed to play nicely with other
 480 * parts of the VM system.
 481 * At each level, we keep a list of pages, which are heads of continuous
 482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
 483 * order is recorded in page_private(page) field.
 484 * So when we are allocating or freeing one, we can derive the state of the
 485 * other.  That is, if we allocate a small block, and both were   
 486 * free, the remainder of the region must be split into blocks.   
 487 * If a block is freed, and its buddy is also free, then this
 488 * triggers coalescing into a block of larger size.            
 489 *
 490 * -- wli
 491 */
 492
 493static inline void __free_one_page(struct page *page,
 
 494		struct zone *zone, unsigned int order,
 495		int migratetype)
 496{
 497	unsigned long page_idx;
 498	unsigned long combined_idx;
 499	unsigned long uninitialized_var(buddy_idx);
 500	struct page *buddy;
 
 501
 502	if (unlikely(PageCompound(page)))
 503		if (unlikely(destroy_compound_page(page, order)))
 504			return;
 505
 506	VM_BUG_ON(migratetype == -1);
 
 
 507
 508	page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 
 509
 510	VM_BUG_ON(page_idx & ((1 << order) - 1));
 511	VM_BUG_ON(bad_range(zone, page));
 
 
 
 
 512
 513	while (order < MAX_ORDER-1) {
 514		buddy_idx = __find_buddy_index(page_idx, order);
 515		buddy = page + (buddy_idx - page_idx);
 516		if (!page_is_buddy(page, buddy, order))
 517			break;
 
 
 
 
 
 
 
 518
 519		/* Our buddy is free, merge with it and move up one order. */
 520		list_del(&buddy->lru);
 521		zone->free_area[order].nr_free--;
 522		rmv_page_order(buddy);
 523		combined_idx = buddy_idx & page_idx;
 524		page = page + (combined_idx - page_idx);
 525		page_idx = combined_idx;
 
 
 
 
 
 
 
 
 
 
 526		order++;
 527	}
 528	set_page_order(page, order);
 529
 530	/*
 531	 * If this is not the largest possible page, check if the buddy
 532	 * of the next-highest order is free. If it is, it's possible
 533	 * that pages are being freed that will coalesce soon. In case,
 534	 * that is happening, add the free page to the tail of the list
 535	 * so it's less likely to be used soon and more likely to be merged
 536	 * as a higher order page
 537	 */
 538	if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
 539		struct page *higher_page, *higher_buddy;
 540		combined_idx = buddy_idx & page_idx;
 541		higher_page = page + (combined_idx - page_idx);
 542		buddy_idx = __find_buddy_index(combined_idx, order + 1);
 543		higher_buddy = page + (buddy_idx - combined_idx);
 544		if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
 545			list_add_tail(&page->lru,
 546				&zone->free_area[order].free_list[migratetype]);
 547			goto out;
 548		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 549	}
 550
 551	list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 552out:
 553	zone->free_area[order].nr_free++;
 
 554}
 555
 556/*
 557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
 558 * Page should not be on lru, so no need to fix that up.
 559 * free_pages_check() will verify...
 560 */
 561static inline void free_page_mlock(struct page *page)
 
 562{
 563	__dec_zone_page_state(page, NR_MLOCK);
 564	__count_vm_event(UNEVICTABLE_MLOCKFREED);
 
 
 
 
 
 
 
 
 
 
 
 
 
 565}
 566
 567static inline int free_pages_check(struct page *page)
 568{
 569	if (unlikely(page_mapcount(page) |
 570		(page->mapping != NULL)  |
 571		(atomic_read(&page->_count) != 0) |
 572		(page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
 573		(mem_cgroup_bad_page_check(page)))) {
 574		bad_page(page);
 575		return 1;
 
 
 
 
 
 
 576	}
 577	if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
 578		page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
 579	return 0;
 
 
 
 
 
 
 580}
 581
 582/*
 583 * Frees a number of pages from the PCP lists
 584 * Assumes all pages on list are in same zone, and of same order.
 585 * count is the number of pages to free.
 586 *
 587 * If the zone was previously in an "all pages pinned" state then look to
 588 * see if this freeing clears that state.
 589 *
 590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
 591 * pinned" detection logic.
 592 */
 593static void free_pcppages_bulk(struct zone *zone, int count,
 594					struct per_cpu_pages *pcp)
 595{
 596	int migratetype = 0;
 597	int batch_free = 0;
 598	int to_free = count;
 599
 600	spin_lock(&zone->lock);
 601	zone->all_unreclaimable = 0;
 602	zone->pages_scanned = 0;
 
 603
 604	while (to_free) {
 605		struct page *page;
 606		struct list_head *list;
 
 607
 608		/*
 609		 * Remove pages from lists in a round-robin fashion. A
 610		 * batch_free count is maintained that is incremented when an
 611		 * empty list is encountered.  This is so more pages are freed
 612		 * off fuller lists instead of spinning excessively around empty
 613		 * lists
 614		 */
 615		do {
 616			batch_free++;
 617			if (++migratetype == MIGRATE_PCPTYPES)
 618				migratetype = 0;
 619			list = &pcp->lists[migratetype];
 620		} while (list_empty(list));
 621
 622		/* This is the only non-empty list. Free them all. */
 623		if (batch_free == MIGRATE_PCPTYPES)
 624			batch_free = to_free;
 
 625
 626		do {
 627			page = list_entry(list->prev, struct page, lru);
 628			/* must delete as __free_one_page list manipulates */
 629			list_del(&page->lru);
 630			/* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
 631			__free_one_page(page, zone, 0, page_private(page));
 632			trace_mm_page_pcpu_drain(page, 0, page_private(page));
 633		} while (--to_free && --batch_free && !list_empty(list));
 
 634	}
 635	__mod_zone_page_state(zone, NR_FREE_PAGES, count);
 636	spin_unlock(&zone->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 637}
 638
 639static void free_one_page(struct zone *zone, struct page *page, int order,
 640				int migratetype)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 641{
 642	spin_lock(&zone->lock);
 643	zone->all_unreclaimable = 0;
 644	zone->pages_scanned = 0;
 645
 646	__free_one_page(page, zone, order, migratetype);
 647	__mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 648	spin_unlock(&zone->lock);
 649}
 650
 651static bool free_pages_prepare(struct page *page, unsigned int order)
 652{
 653	int i;
 
 
 
 
 
 
 
 
 
 
 
 654	int bad = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 655
 656	trace_mm_page_free_direct(page, order);
 657	kmemcheck_free_shadow(page, order);
 
 
 
 
 658
 659	if (PageAnon(page))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 660		page->mapping = NULL;
 661	for (i = 0; i < (1 << order); i++)
 662		bad += free_pages_check(page + i);
 663	if (bad)
 664		return false;
 
 
 
 
 
 
 
 665
 666	if (!PageHighMem(page)) {
 667		debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
 
 668		debug_check_no_obj_freed(page_address(page),
 669					   PAGE_SIZE << order);
 670	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 671	arch_free_page(page, order);
 672	kernel_map_pages(page, 1 << order, 0);
 
 673
 674	return true;
 675}
 676
 677static void __free_pages_ok(struct page *page, unsigned int order)
 
 
 
 
 
 
 
 678{
 679	unsigned long flags;
 680	int wasMlocked = __TestClearPageMlocked(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 681
 682	if (!free_pages_prepare(page, order))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 683		return;
 684
 685	local_irq_save(flags);
 686	if (unlikely(wasMlocked))
 687		free_page_mlock(page);
 
 
 
 
 
 
 688	__count_vm_events(PGFREE, 1 << order);
 689	free_one_page(page_zone(page), page, order,
 690					get_pageblock_migratetype(page));
 691	local_irq_restore(flags);
 692}
 693
 694/*
 695 * permit the bootmem allocator to evade page validation on high-order frees
 696 */
 697void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
 698{
 699	if (order == 0) {
 700		__ClearPageReserved(page);
 701		set_page_count(page, 0);
 702		set_page_refcounted(page);
 703		__free_page(page);
 704	} else {
 705		int loop;
 706
 707		prefetchw(page);
 708		for (loop = 0; loop < BITS_PER_LONG; loop++) {
 709			struct page *p = &page[loop];
 
 
 
 
 
 
 
 
 
 
 710
 711			if (loop + 1 < BITS_PER_LONG)
 712				prefetchw(p + 1);
 713			__ClearPageReserved(p);
 714			set_page_count(p, 0);
 715		}
 716
 717		set_page_refcounted(page);
 718		__free_pages(page, order);
 
 
 
 719	}
 
 
 
 
 
 
 720}
 721
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 722
 723/*
 724 * The order of subdivision here is critical for the IO subsystem.
 725 * Please do not alter this order without good reasons and regression
 726 * testing. Specifically, as large blocks of memory are subdivided,
 727 * the order in which smaller blocks are delivered depends on the order
 728 * they're subdivided in this function. This is the primary factor
 729 * influencing the order in which pages are delivered to the IO
 730 * subsystem according to empirical testing, and this is also justified
 731 * by considering the behavior of a buddy system containing a single
 732 * large block of memory acted on by a series of small allocations.
 733 * This behavior is a critical factor in sglist merging's success.
 734 *
 735 * -- wli
 736 */
 737static inline void expand(struct zone *zone, struct page *page,
 738	int low, int high, struct free_area *area,
 739	int migratetype)
 740{
 741	unsigned long size = 1 << high;
 742
 743	while (high > low) {
 744		area--;
 745		high--;
 746		size >>= 1;
 747		VM_BUG_ON(bad_range(zone, &page[size]));
 748		list_add(&page[size].lru, &area->free_list[migratetype]);
 749		area->nr_free++;
 750		set_page_order(&page[size], high);
 
 
 
 
 
 
 
 
 
 751	}
 752}
 753
 
 
 
 
 
 
 
 
 
 
 
 
 754/*
 755 * This page is about to be returned from the page allocator
 756 */
 757static inline int check_new_page(struct page *page)
 758{
 759	if (unlikely(page_mapcount(page) |
 760		(page->mapping != NULL)  |
 761		(atomic_read(&page->_count) != 0)  |
 762		(page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
 763		(mem_cgroup_bad_page_check(page)))) {
 764		bad_page(page);
 765		return 1;
 766	}
 767	return 0;
 768}
 769
 770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
 771{
 772	int i;
 
 
 773
 774	for (i = 0; i < (1 << order); i++) {
 775		struct page *p = page + i;
 776		if (unlikely(check_new_page(p)))
 777			return 1;
 778	}
 779
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 780	set_page_private(page, 0);
 781	set_page_refcounted(page);
 782
 783	arch_alloc_page(page, order);
 784	kernel_map_pages(page, 1 << order, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 785
 786	if (gfp_flags & __GFP_ZERO)
 787		prep_zero_page(page, order, gfp_flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 788
 789	if (order && (gfp_flags & __GFP_COMP))
 790		prep_compound_page(page, order);
 791
 792	return 0;
 
 
 
 
 
 
 
 
 
 793}
 794
 795/*
 796 * Go through the free lists for the given migratetype and remove
 797 * the smallest available page from the freelists
 798 */
 799static inline
 800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
 801						int migratetype)
 802{
 803	unsigned int current_order;
 804	struct free_area * area;
 805	struct page *page;
 806
 807	/* Find a page of the appropriate size in the preferred list */
 808	for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 809		area = &(zone->free_area[current_order]);
 810		if (list_empty(&area->free_list[migratetype]))
 
 811			continue;
 812
 813		page = list_entry(area->free_list[migratetype].next,
 814							struct page, lru);
 815		list_del(&page->lru);
 816		rmv_page_order(page);
 817		area->nr_free--;
 818		expand(zone, page, order, current_order, area, migratetype);
 819		return page;
 820	}
 821
 822	return NULL;
 823}
 824
 825
 826/*
 827 * This array describes the order lists are fallen back to when
 828 * the free lists for the desirable migrate type are depleted
 
 
 829 */
 830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
 831	[MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 832	[MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_RESERVE },
 833	[MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
 834	[MIGRATE_RESERVE]     = { MIGRATE_RESERVE,     MIGRATE_RESERVE,   MIGRATE_RESERVE }, /* Never used */
 835};
 836
 
 
 
 
 
 
 
 
 
 
 
 837/*
 838 * Move the free pages in a range to the free lists of the requested type.
 839 * Note that start_page and end_pages are not aligned on a pageblock
 840 * boundary. If alignment is required, use move_freepages_block()
 841 */
 842static int move_freepages(struct zone *zone,
 843			  struct page *start_page, struct page *end_page,
 844			  int migratetype)
 845{
 846	struct page *page;
 847	unsigned long order;
 
 848	int pages_moved = 0;
 849
 850#ifndef CONFIG_HOLES_IN_ZONE
 851	/*
 852	 * page_zone is not safe to call in this context when
 853	 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
 854	 * anyway as we check zone boundaries in move_freepages_block().
 855	 * Remove at a later date when no bug reports exist related to
 856	 * grouping pages by mobility
 857	 */
 858	BUG_ON(page_zone(start_page) != page_zone(end_page));
 859#endif
 860
 861	for (page = start_page; page <= end_page;) {
 862		/* Make sure we are not inadvertently changing nodes */
 863		VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
 864
 865		if (!pfn_valid_within(page_to_pfn(page))) {
 866			page++;
 867			continue;
 868		}
 869
 870		if (!PageBuddy(page)) {
 871			page++;
 
 
 
 
 
 
 
 
 872			continue;
 873		}
 874
 875		order = page_order(page);
 876		list_move(&page->lru,
 877			  &zone->free_area[order].free_list[migratetype]);
 878		page += 1 << order;
 
 
 
 879		pages_moved += 1 << order;
 880	}
 881
 882	return pages_moved;
 883}
 884
 885static int move_freepages_block(struct zone *zone, struct page *page,
 886				int migratetype)
 887{
 888	unsigned long start_pfn, end_pfn;
 889	struct page *start_page, *end_page;
 890
 891	start_pfn = page_to_pfn(page);
 892	start_pfn = start_pfn & ~(pageblock_nr_pages-1);
 893	start_page = pfn_to_page(start_pfn);
 894	end_page = start_page + pageblock_nr_pages - 1;
 895	end_pfn = start_pfn + pageblock_nr_pages - 1;
 
 896
 897	/* Do not cross zone boundaries */
 898	if (start_pfn < zone->zone_start_pfn)
 899		start_page = page;
 900	if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
 901		return 0;
 902
 903	return move_freepages(zone, start_page, end_page, migratetype);
 
 904}
 905
 906static void change_pageblock_range(struct page *pageblock_page,
 907					int start_order, int migratetype)
 908{
 909	int nr_pageblocks = 1 << (start_order - pageblock_order);
 910
 911	while (nr_pageblocks--) {
 912		set_pageblock_migratetype(pageblock_page, migratetype);
 913		pageblock_page += pageblock_nr_pages;
 914	}
 915}
 916
 917/* Remove an element from the buddy allocator from the fallback list */
 918static inline struct page *
 919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 920{
 921	struct free_area * area;
 922	int current_order;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 923	struct page *page;
 924	int migratetype, i;
 
 925
 926	/* Find the largest possible block of pages in the other list */
 927	for (current_order = MAX_ORDER-1; current_order >= order;
 928						--current_order) {
 929		for (i = 0; i < MIGRATE_TYPES - 1; i++) {
 930			migratetype = fallbacks[start_migratetype][i];
 
 
 
 
 931
 932			/* MIGRATE_RESERVE handled later if necessary */
 933			if (migratetype == MIGRATE_RESERVE)
 934				continue;
 935
 936			area = &(zone->free_area[current_order]);
 937			if (list_empty(&area->free_list[migratetype]))
 938				continue;
 939
 940			page = list_entry(area->free_list[migratetype].next,
 941					struct page, lru);
 942			area->nr_free--;
 943
 944			/*
 945			 * If breaking a large block of pages, move all free
 946			 * pages to the preferred allocation list. If falling
 947			 * back for a reclaimable kernel allocation, be more
 948			 * aggressive about taking ownership of free pages
 
 949			 */
 950			if (unlikely(current_order >= (pageblock_order >> 1)) ||
 951					start_migratetype == MIGRATE_RECLAIMABLE ||
 952					page_group_by_mobility_disabled) {
 953				unsigned long pages;
 954				pages = move_freepages_block(zone, page,
 955								start_migratetype);
 956
 957				/* Claim the whole block if over half of it is free */
 958				if (pages >= (1 << (pageblock_order-1)) ||
 959						page_group_by_mobility_disabled)
 960					set_pageblock_migratetype(page,
 961								start_migratetype);
 962
 963				migratetype = start_migratetype;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 964			}
 
 
 
 965
 966			/* Remove the page from the freelists */
 967			list_del(&page->lru);
 968			rmv_page_order(page);
 969
 970			/* Take ownership for orders >= pageblock_order */
 971			if (current_order >= pageblock_order)
 972				change_pageblock_range(page, current_order,
 973							start_migratetype);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 974
 975			expand(zone, page, order, current_order, area, migratetype);
 
 
 
 
 
 
 976
 977			trace_mm_page_alloc_extfrag(page, order, current_order,
 978				start_migratetype, migratetype);
 
 
 
 
 
 
 
 
 
 
 979
 980			return page;
 981		}
 
 
 
 
 
 
 
 
 
 
 
 982	}
 983
 984	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 985}
 986
 987/*
 988 * Do the hard work of removing an element from the buddy allocator.
 989 * Call me with the zone->lock already held.
 990 */
 991static struct page *__rmqueue(struct zone *zone, unsigned int order,
 992						int migratetype)
 
 993{
 994	struct page *page;
 995
 996retry_reserve:
 997	page = __rmqueue_smallest(zone, order, migratetype);
 998
 999	if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000		page = __rmqueue_fallback(zone, order, migratetype);
1001
1002		/*
1003		 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004		 * is used because __rmqueue_smallest is an inline function
1005		 * and we want just one call site
1006		 */
1007		if (!page) {
1008			migratetype = MIGRATE_RESERVE;
1009			goto retry_reserve;
 
 
 
1010		}
1011	}
1012
1013	trace_mm_page_alloc_zone_locked(page, order, migratetype);
 
 
 
 
 
 
 
 
1014	return page;
1015}
1016
1017/* 
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency.  Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1021 */
1022static int rmqueue_bulk(struct zone *zone, unsigned int order, 
1023			unsigned long count, struct list_head *list,
1024			int migratetype, int cold)
1025{
 
1026	int i;
1027	
1028	spin_lock(&zone->lock);
1029	for (i = 0; i < count; ++i) {
1030		struct page *page = __rmqueue(zone, order, migratetype);
 
1031		if (unlikely(page == NULL))
1032			break;
1033
1034		/*
1035		 * Split buddy pages returned by expand() are received here
1036		 * in physical page order. The page is added to the callers and
1037		 * list and the list head then moves forward. From the callers
1038		 * perspective, the linked list is ordered by page number in
1039		 * some conditions. This is useful for IO devices that can
1040		 * merge IO requests if the physical pages are ordered
1041		 * properly.
 
1042		 */
1043		if (likely(cold == 0))
1044			list_add(&page->lru, list);
1045		else
1046			list_add_tail(&page->lru, list);
1047		set_page_private(page, migratetype);
1048		list = &page->lru;
1049	}
 
1050	__mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051	spin_unlock(&zone->lock);
 
1052	return i;
1053}
1054
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1055#ifdef CONFIG_NUMA
1056/*
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1059 * expired.
1060 *
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1063 */
1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1065{
1066	unsigned long flags;
1067	int to_drain;
1068
1069	local_irq_save(flags);
1070	if (pcp->count >= pcp->batch)
1071		to_drain = pcp->batch;
1072	else
1073		to_drain = pcp->count;
1074	free_pcppages_bulk(zone, to_drain, pcp);
1075	pcp->count -= to_drain;
1076	local_irq_restore(flags);
1077}
1078#endif
1079
1080/*
1081 * Drain pages of the indicated processor.
1082 *
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1085 * is not online.
 
 
 
 
 
 
 
 
 
 
 
1086 */
1087static void drain_pages(unsigned int cpu)
1088{
1089	unsigned long flags;
1090	struct zone *zone;
1091
1092	for_each_populated_zone(zone) {
1093		struct per_cpu_pageset *pset;
1094		struct per_cpu_pages *pcp;
1095
1096		local_irq_save(flags);
1097		pset = per_cpu_ptr(zone->pageset, cpu);
1098
1099		pcp = &pset->pcp;
1100		if (pcp->count) {
1101			free_pcppages_bulk(zone, pcp->count, pcp);
1102			pcp->count = 0;
1103		}
1104		local_irq_restore(flags);
1105	}
1106}
1107
1108/*
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1110 */
1111void drain_local_pages(void *arg)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1112{
1113	drain_pages(smp_processor_id());
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1114}
1115
1116/*
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
 
 
1118 */
1119void drain_all_pages(void)
1120{
1121	on_each_cpu(drain_local_pages, NULL, 1);
1122}
1123
1124#ifdef CONFIG_HIBERNATION
 
 
 
 
 
 
1125
1126void mark_free_pages(struct zone *zone)
 
 
 
 
 
1127{
1128	unsigned long pfn, max_zone_pfn;
1129	unsigned long flags;
1130	int order, t;
1131	struct list_head *curr;
1132
1133	if (!zone->spanned_pages)
1134		return;
 
1135
1136	spin_lock_irqsave(&zone->lock, flags);
 
 
1137
1138	max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139	for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140		if (pfn_valid(pfn)) {
1141			struct page *page = pfn_to_page(pfn);
1142
1143			if (!swsusp_page_is_forbidden(page))
1144				swsusp_unset_page_free(page);
1145		}
1146
1147	for_each_migratetype_order(order, t) {
1148		list_for_each(curr, &zone->free_area[order].free_list[t]) {
1149			unsigned long i;
1150
1151			pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152			for (i = 0; i < (1UL << order); i++)
1153				swsusp_set_page_free(pfn_to_page(pfn + i));
1154		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1155	}
1156	spin_unlock_irqrestore(&zone->lock, flags);
1157}
1158#endif /* CONFIG_PM */
1159
1160/*
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1163 */
1164void free_hot_cold_page(struct page *page, int cold)
1165{
1166	struct zone *zone = page_zone(page);
1167	struct per_cpu_pages *pcp;
1168	unsigned long flags;
1169	int migratetype;
1170	int wasMlocked = __TestClearPageMlocked(page);
1171
1172	if (!free_pages_prepare(page, 0))
1173		return;
1174
1175	migratetype = get_pageblock_migratetype(page);
1176	set_page_private(page, migratetype);
1177	local_irq_save(flags);
1178	if (unlikely(wasMlocked))
1179		free_page_mlock(page);
1180	__count_vm_event(PGFREE);
1181
1182	/*
1183	 * We only track unmovable, reclaimable and movable on pcp lists.
1184	 * Free ISOLATE pages back to the allocator because they are being
1185	 * offlined but treat RESERVE as movable pages so we can get those
1186	 * areas back if necessary. Otherwise, we may have to free
1187	 * excessively into the page allocator
1188	 */
1189	if (migratetype >= MIGRATE_PCPTYPES) {
1190		if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191			free_one_page(zone, page, 0, migratetype);
1192			goto out;
 
1193		}
1194		migratetype = MIGRATE_MOVABLE;
1195	}
1196
1197	pcp = &this_cpu_ptr(zone->pageset)->pcp;
1198	if (cold)
1199		list_add_tail(&page->lru, &pcp->lists[migratetype]);
1200	else
1201		list_add(&page->lru, &pcp->lists[migratetype]);
1202	pcp->count++;
1203	if (pcp->count >= pcp->high) {
1204		free_pcppages_bulk(zone, pcp->batch, pcp);
1205		pcp->count -= pcp->batch;
1206	}
 
 
1207
1208out:
1209	local_irq_restore(flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1210}
1211
1212/*
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1216 *
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1219 */
1220void split_page(struct page *page, unsigned int order)
1221{
1222	int i;
1223
1224	VM_BUG_ON(PageCompound(page));
1225	VM_BUG_ON(!page_count(page));
1226
1227#ifdef CONFIG_KMEMCHECK
1228	/*
1229	 * Split shadow pages too, because free(page[0]) would
1230	 * otherwise free the whole shadow.
1231	 */
1232	if (kmemcheck_page_is_tracked(page))
1233		split_page(virt_to_page(page[0].shadow), order);
1234#endif
1235
1236	for (i = 1; i < (1 << order); i++)
1237		set_page_refcounted(page + i);
 
 
1238}
 
1239
1240/*
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * are enabled.
1246 *
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1249 */
1250int split_free_page(struct page *page)
1251{
1252	unsigned int order;
1253	unsigned long watermark;
1254	struct zone *zone;
1255
1256	BUG_ON(!PageBuddy(page));
1257
1258	zone = page_zone(page);
1259	order = page_order(page);
1260
1261	/* Obey watermarks as if the page was being allocated */
1262	watermark = low_wmark_pages(zone) + (1 << order);
1263	if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1264		return 0;
 
 
 
 
 
 
 
1265
1266	/* Remove page from free list */
1267	list_del(&page->lru);
1268	zone->free_area[order].nr_free--;
1269	rmv_page_order(page);
1270	__mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271
1272	/* Split into individual pages */
1273	set_page_refcounted(page);
1274	split_page(page, order);
1275
 
 
 
 
1276	if (order >= pageblock_order - 1) {
1277		struct page *endpage = page + (1 << order) - 1;
1278		for (; page < endpage; page += pageblock_nr_pages)
1279			set_pageblock_migratetype(page, MIGRATE_MOVABLE);
 
 
 
 
 
 
 
 
1280	}
1281
1282	return 1 << order;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1283}
1284
1285/*
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
1287 * we cheat by calling it from here, in the order > 0 path.  Saves a branch
1288 * or two.
1289 */
1290static inline
1291struct page *buffered_rmqueue(struct zone *preferred_zone,
1292			struct zone *zone, int order, gfp_t gfp_flags,
1293			int migratetype)
1294{
1295	unsigned long flags;
1296	struct page *page;
1297	int cold = !!(gfp_flags & __GFP_COLD);
1298
1299again:
1300	if (likely(order == 0)) {
1301		struct per_cpu_pages *pcp;
1302		struct list_head *list;
1303
1304		local_irq_save(flags);
1305		pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306		list = &pcp->lists[migratetype];
1307		if (list_empty(list)) {
1308			pcp->count += rmqueue_bulk(zone, 0,
1309					pcp->batch, list,
1310					migratetype, cold);
1311			if (unlikely(list_empty(list)))
1312				goto failed;
1313		}
1314
1315		if (cold)
1316			page = list_entry(list->prev, struct page, lru);
1317		else
1318			page = list_entry(list->next, struct page, lru);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1319
1320		list_del(&page->lru);
1321		pcp->count--;
1322	} else {
1323		if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324			/*
1325			 * __GFP_NOFAIL is not to be used in new code.
1326			 *
1327			 * All __GFP_NOFAIL callers should be fixed so that they
1328			 * properly detect and handle allocation failures.
1329			 *
1330			 * We most definitely don't want callers attempting to
1331			 * allocate greater than order-1 page units with
1332			 * __GFP_NOFAIL.
1333			 */
1334			WARN_ON_ONCE(order > 1);
 
 
 
 
 
 
1335		}
1336		spin_lock_irqsave(&zone->lock, flags);
1337		page = __rmqueue(zone, order, migratetype);
1338		spin_unlock(&zone->lock);
1339		if (!page)
1340			goto failed;
1341		__mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1342	}
1343
1344	__count_zone_vm_events(PGALLOC, zone, 1 << order);
1345	zone_statistics(preferred_zone, zone, gfp_flags);
1346	local_irq_restore(flags);
1347
1348	VM_BUG_ON(bad_range(zone, page));
1349	if (prep_new_page(page, order, gfp_flags))
1350		goto again;
1351	return page;
1352
1353failed:
1354	local_irq_restore(flags);
1355	return NULL;
1356}
1357
1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359#define ALLOC_WMARK_MIN		WMARK_MIN
1360#define ALLOC_WMARK_LOW		WMARK_LOW
1361#define ALLOC_WMARK_HIGH	WMARK_HIGH
1362#define ALLOC_NO_WATERMARKS	0x04 /* don't check watermarks at all */
1363
1364/* Mask to get the watermark bits */
1365#define ALLOC_WMARK_MASK	(ALLOC_NO_WATERMARKS-1)
1366
1367#define ALLOC_HARDER		0x10 /* try to alloc harder */
1368#define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
1369#define ALLOC_CPUSET		0x40 /* check for correct cpuset */
1370
1371#ifdef CONFIG_FAIL_PAGE_ALLOC
1372
1373static struct {
1374	struct fault_attr attr;
1375
1376	u32 ignore_gfp_highmem;
1377	u32 ignore_gfp_wait;
1378	u32 min_order;
1379} fail_page_alloc = {
1380	.attr = FAULT_ATTR_INITIALIZER,
1381	.ignore_gfp_wait = 1,
1382	.ignore_gfp_highmem = 1,
1383	.min_order = 1,
1384};
1385
1386static int __init setup_fail_page_alloc(char *str)
1387{
1388	return setup_fault_attr(&fail_page_alloc.attr, str);
1389}
1390__setup("fail_page_alloc=", setup_fail_page_alloc);
1391
1392static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393{
1394	if (order < fail_page_alloc.min_order)
1395		return 0;
1396	if (gfp_mask & __GFP_NOFAIL)
1397		return 0;
1398	if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399		return 0;
1400	if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1401		return 0;
1402
1403	return should_fail(&fail_page_alloc.attr, 1 << order);
1404}
 
1405
1406#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 
 
 
1407
1408static int __init fail_page_alloc_debugfs(void)
1409{
1410	mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1411	struct dentry *dir;
1412
1413	dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1414					&fail_page_alloc.attr);
1415	if (IS_ERR(dir))
1416		return PTR_ERR(dir);
1417
1418	if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1419				&fail_page_alloc.ignore_gfp_wait))
1420		goto fail;
1421	if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1422				&fail_page_alloc.ignore_gfp_highmem))
1423		goto fail;
1424	if (!debugfs_create_u32("min-order", mode, dir,
1425				&fail_page_alloc.min_order))
1426		goto fail;
1427
1428	return 0;
1429fail:
1430	debugfs_remove_recursive(dir);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1431
1432	return -ENOMEM;
1433}
1434
1435late_initcall(fail_page_alloc_debugfs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1436
1437#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
 
 
 
1438
1439#else /* CONFIG_FAIL_PAGE_ALLOC */
 
1440
1441static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
 
 
 
1442{
1443	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1444}
1445
1446#endif /* CONFIG_FAIL_PAGE_ALLOC */
 
 
 
1447
1448/*
1449 * Return true if free pages are above 'mark'. This takes into account the order
1450 * of the allocation.
 
 
1451 */
1452static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1453		      int classzone_idx, int alloc_flags, long free_pages)
 
 
 
 
1454{
1455	/* free_pages my go negative - that's OK */
1456	long min = mark;
1457	int o;
1458
1459	free_pages -= (1 << order) + 1;
1460	if (alloc_flags & ALLOC_HIGH)
1461		min -= min / 2;
1462	if (alloc_flags & ALLOC_HARDER)
1463		min -= min / 4;
1464
1465	if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1466		return false;
1467	for (o = 0; o < order; o++) {
1468		/* At the next order, this order's pages become unavailable */
1469		free_pages -= z->free_area[o].nr_free << o;
 
1470
1471		/* Require fewer higher order pages to be free */
1472		min >>= 1;
1473
1474		if (free_pages <= min)
1475			return false;
 
 
 
 
1476	}
1477	return true;
 
 
1478}
1479
1480bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1481		      int classzone_idx, int alloc_flags)
1482{
1483	return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1484					zone_page_state(z, NR_FREE_PAGES));
1485}
 
1486
1487bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1488		      int classzone_idx, int alloc_flags)
1489{
1490	long free_pages = zone_page_state(z, NR_FREE_PAGES);
1491
1492	if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1493		free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
 
 
 
 
 
1494
1495	return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1496								free_pages);
 
 
 
 
 
 
 
 
1497}
1498
1499#ifdef CONFIG_NUMA
1500/*
1501 * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1502 * skip over zones that are not allowed by the cpuset, or that have
1503 * been recently (in last second) found to be nearly full.  See further
1504 * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1505 * that have to skip over a lot of full or unallowed zones.
1506 *
1507 * If the zonelist cache is present in the passed in zonelist, then
1508 * returns a pointer to the allowed node mask (either the current
1509 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1510 *
1511 * If the zonelist cache is not available for this zonelist, does
1512 * nothing and returns NULL.
1513 *
1514 * If the fullzones BITMAP in the zonelist cache is stale (more than
1515 * a second since last zap'd) then we zap it out (clear its bits.)
1516 *
1517 * We hold off even calling zlc_setup, until after we've checked the
1518 * first zone in the zonelist, on the theory that most allocations will
1519 * be satisfied from that first zone, so best to examine that zone as
1520 * quickly as we can.
1521 */
1522static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
 
 
1523{
1524	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1525	nodemask_t *allowednodes;	/* zonelist_cache approximation */
1526
1527	zlc = zonelist->zlcache_ptr;
1528	if (!zlc)
1529		return NULL;
1530
1531	if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1532		bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1533		zlc->last_full_zap = jiffies;
1534	}
1535
1536	allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1537					&cpuset_current_mems_allowed :
1538					&node_states[N_HIGH_MEMORY];
1539	return allowednodes;
1540}
1541
1542/*
1543 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1544 * if it is worth looking at further for free memory:
1545 *  1) Check that the zone isn't thought to be full (doesn't have its
1546 *     bit set in the zonelist_cache fullzones BITMAP).
1547 *  2) Check that the zones node (obtained from the zonelist_cache
1548 *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1549 * Return true (non-zero) if zone is worth looking at further, or
1550 * else return false (zero) if it is not.
1551 *
1552 * This check -ignores- the distinction between various watermarks,
1553 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1554 * found to be full for any variation of these watermarks, it will
1555 * be considered full for up to one second by all requests, unless
1556 * we are so low on memory on all allowed nodes that we are forced
1557 * into the second scan of the zonelist.
1558 *
1559 * In the second scan we ignore this zonelist cache and exactly
1560 * apply the watermarks to all zones, even it is slower to do so.
1561 * We are low on memory in the second scan, and should leave no stone
1562 * unturned looking for a free page.
1563 */
1564static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1565						nodemask_t *allowednodes)
1566{
1567	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1568	int i;				/* index of *z in zonelist zones */
1569	int n;				/* node that zone *z is on */
1570
1571	zlc = zonelist->zlcache_ptr;
1572	if (!zlc)
1573		return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1574
1575	i = z - zonelist->_zonerefs;
1576	n = zlc->z_to_n[i];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1577
1578	/* This zone is worth trying if it is allowed but not full */
1579	return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
 
 
 
 
 
 
 
 
 
 
1580}
1581
1582/*
1583 * Given 'z' scanning a zonelist, set the corresponding bit in
1584 * zlc->fullzones, so that subsequent attempts to allocate a page
1585 * from that zone don't waste time re-examining it.
1586 */
1587static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1588{
1589	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1590	int i;				/* index of *z in zonelist zones */
 
1591
1592	zlc = zonelist->zlcache_ptr;
1593	if (!zlc)
1594		return;
 
 
1595
1596	i = z - zonelist->_zonerefs;
1597
1598	set_bit(i, zlc->fullzones);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1599}
1600
1601/*
1602 * clear all zones full, called after direct reclaim makes progress so that
1603 * a zone that was recently full is not skipped over for up to a second
1604 */
1605static void zlc_clear_zones_full(struct zonelist *zonelist)
1606{
1607	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1608
1609	zlc = zonelist->zlcache_ptr;
1610	if (!zlc)
1611		return;
1612
1613	bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
 
1614}
1615
1616#else	/* CONFIG_NUMA */
 
1617
1618static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1619{
1620	return NULL;
 
1621}
1622
1623static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1624				nodemask_t *allowednodes)
1625{
1626	return 1;
1627}
 
1628
1629static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
 
 
 
 
 
 
 
 
 
1630{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1631}
1632
1633static void zlc_clear_zones_full(struct zonelist *zonelist)
1634{
 
 
 
 
 
 
 
1635}
1636#endif	/* CONFIG_NUMA */
1637
1638/*
1639 * get_page_from_freelist goes through the zonelist trying to allocate
1640 * a page.
1641 */
1642static struct page *
1643get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1644		struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1645		struct zone *preferred_zone, int migratetype)
1646{
1647	struct zoneref *z;
1648	struct page *page = NULL;
1649	int classzone_idx;
1650	struct zone *zone;
1651	nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1652	int zlc_active = 0;		/* set if using zonelist_cache */
1653	int did_zlc_setup = 0;		/* just call zlc_setup() one time */
1654
1655	classzone_idx = zone_idx(preferred_zone);
1656zonelist_scan:
1657	/*
1658	 * Scan zonelist, looking for a zone with enough free.
1659	 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1660	 */
1661	for_each_zone_zonelist_nodemask(zone, z, zonelist,
1662						high_zoneidx, nodemask) {
1663		if (NUMA_BUILD && zlc_active &&
1664			!zlc_zone_worth_trying(zonelist, z, allowednodes))
1665				continue;
1666		if ((alloc_flags & ALLOC_CPUSET) &&
1667			!cpuset_zone_allowed_softwall(zone, gfp_mask))
 
 
 
1668				continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1669
1670		BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1671		if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1672			unsigned long mark;
1673			int ret;
1674
1675			mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1676			if (zone_watermark_ok(zone, order, mark,
1677				    classzone_idx, alloc_flags))
1678				goto try_this_zone;
1679
1680			if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681				/*
1682				 * we do zlc_setup if there are multiple nodes
1683				 * and before considering the first zone allowed
1684				 * by the cpuset.
1685				 */
1686				allowednodes = zlc_setup(zonelist, alloc_flags);
1687				zlc_active = 1;
1688				did_zlc_setup = 1;
1689			}
 
1690
1691			if (zone_reclaim_mode == 0)
1692				goto this_zone_full;
 
 
 
 
 
 
 
1693
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1694			/*
1695			 * As we may have just activated ZLC, check if the first
1696			 * eligible zone has failed zone_reclaim recently.
1697			 */
1698			if (NUMA_BUILD && zlc_active &&
1699				!zlc_zone_worth_trying(zonelist, z, allowednodes))
 
 
 
 
 
 
 
 
 
 
1700				continue;
1701
1702			ret = zone_reclaim(zone, gfp_mask, order);
1703			switch (ret) {
1704			case ZONE_RECLAIM_NOSCAN:
1705				/* did not scan */
1706				continue;
1707			case ZONE_RECLAIM_FULL:
1708				/* scanned but unreclaimable */
1709				continue;
1710			default:
1711				/* did we reclaim enough */
1712				if (!zone_watermark_ok(zone, order, mark,
1713						classzone_idx, alloc_flags))
1714					goto this_zone_full;
 
 
1715			}
1716		}
1717
1718try_this_zone:
1719		page = buffered_rmqueue(preferred_zone, zone, order,
1720						gfp_mask, migratetype);
1721		if (page)
1722			break;
1723this_zone_full:
1724		if (NUMA_BUILD)
1725			zlc_mark_zone_full(zonelist, z);
1726	}
1727
1728	if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1729		/* Disable zlc cache for second zonelist scan */
1730		zlc_active = 0;
1731		goto zonelist_scan;
1732	}
1733	return page;
1734}
1735
1736/*
1737 * Large machines with many possible nodes should not always dump per-node
1738 * meminfo in irq context.
1739 */
1740static inline bool should_suppress_show_mem(void)
1741{
1742	bool ret = false;
1743
1744#if NODES_SHIFT > 8
1745	ret = in_interrupt();
 
 
 
 
1746#endif
1747	return ret;
1748}
 
 
 
 
 
 
 
 
 
1749
1750static DEFINE_RATELIMIT_STATE(nopage_rs,
1751		DEFAULT_RATELIMIT_INTERVAL,
1752		DEFAULT_RATELIMIT_BURST);
1753
1754void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1755{
1756	va_list args;
1757	unsigned int filter = SHOW_MEM_FILTER_NODES;
1758
1759	if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1760		return;
1761
1762	/*
1763	 * This documents exceptions given to allocations in certain
1764	 * contexts that are allowed to allocate outside current's set
1765	 * of allowed nodes.
1766	 */
1767	if (!(gfp_mask & __GFP_NOMEMALLOC))
1768		if (test_thread_flag(TIF_MEMDIE) ||
1769		    (current->flags & (PF_MEMALLOC | PF_EXITING)))
1770			filter &= ~SHOW_MEM_FILTER_NODES;
1771	if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1772		filter &= ~SHOW_MEM_FILTER_NODES;
1773
1774	if (fmt) {
1775		printk(KERN_WARNING);
1776		va_start(args, fmt);
1777		vprintk(fmt, args);
1778		va_end(args);
1779	}
 
 
1780
1781	pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1782		   current->comm, order, gfp_mask);
 
 
 
 
 
 
 
 
 
 
1783
 
 
1784	dump_stack();
1785	if (!should_suppress_show_mem())
1786		show_mem(filter);
1787}
1788
1789static inline int
1790should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1791				unsigned long pages_reclaimed)
 
1792{
1793	/* Do not loop if specifically requested */
1794	if (gfp_mask & __GFP_NORETRY)
1795		return 0;
1796
1797	/*
1798	 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1799	 * means __GFP_NOFAIL, but that may not be true in other
1800	 * implementations.
1801	 */
1802	if (order <= PAGE_ALLOC_COSTLY_ORDER)
1803		return 1;
1804
1805	/*
1806	 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1807	 * specified, then we retry until we no longer reclaim any pages
1808	 * (above), or we've reclaimed an order of pages at least as
1809	 * large as the allocation's order. In both cases, if the
1810	 * allocation still fails, we stop retrying.
1811	 */
1812	if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1813		return 1;
1814
 
 
1815	/*
1816	 * Don't let big-order allocations loop unless the caller
1817	 * explicitly requests that.
1818	 */
1819	if (gfp_mask & __GFP_NOFAIL)
1820		return 1;
 
1821
1822	return 0;
1823}
1824
1825static inline struct page *
1826__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1827	struct zonelist *zonelist, enum zone_type high_zoneidx,
1828	nodemask_t *nodemask, struct zone *preferred_zone,
1829	int migratetype)
1830{
 
 
 
 
 
 
 
1831	struct page *page;
1832
1833	/* Acquire the OOM killer lock for the zones in zonelist */
1834	if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
 
 
 
 
 
 
1835		schedule_timeout_uninterruptible(1);
1836		return NULL;
1837	}
1838
1839	/*
1840	 * Go through the zonelist yet one more time, keep very high watermark
1841	 * here, this is only to catch a parallel oom killing, we must fail if
1842	 * we're still under heavy pressure.
1843	 */
1844	page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1845		order, zonelist, high_zoneidx,
1846		ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1847		preferred_zone, migratetype);
 
1848	if (page)
1849		goto out;
1850
1851	if (!(gfp_mask & __GFP_NOFAIL)) {
1852		/* The OOM killer will not help higher order allocs */
1853		if (order > PAGE_ALLOC_COSTLY_ORDER)
1854			goto out;
1855		/* The OOM killer does not needlessly kill tasks for lowmem */
1856		if (high_zoneidx < ZONE_NORMAL)
1857			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1858		/*
1859		 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1860		 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1861		 * The caller should handle page allocation failure by itself if
1862		 * it specifies __GFP_THISNODE.
1863		 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1864		 */
1865		if (gfp_mask & __GFP_THISNODE)
1866			goto out;
 
1867	}
1868	/* Exhausted what can be done so it's blamo time */
1869	out_of_memory(zonelist, gfp_mask, order, nodemask);
1870
1871out:
1872	clear_zonelist_oom(zonelist, gfp_mask);
1873	return page;
1874}
1875
 
 
 
 
 
 
1876#ifdef CONFIG_COMPACTION
1877/* Try memory compaction for high-order allocations before reclaim */
1878static struct page *
1879__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1880	struct zonelist *zonelist, enum zone_type high_zoneidx,
1881	nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1882	int migratetype, unsigned long *did_some_progress,
1883	bool sync_migration)
1884{
1885	struct page *page;
 
 
1886
1887	if (!order || compaction_deferred(preferred_zone))
1888		return NULL;
1889
1890	current->flags |= PF_MEMALLOC;
1891	*did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1892						nodemask, sync_migration);
1893	current->flags &= ~PF_MEMALLOC;
1894	if (*did_some_progress != COMPACT_SKIPPED) {
1895
1896		/* Page migration frees to the PCP lists but we want merging */
1897		drain_pages(get_cpu());
1898		put_cpu();
1899
1900		page = get_page_from_freelist(gfp_mask, nodemask,
1901				order, zonelist, high_zoneidx,
1902				alloc_flags, preferred_zone,
1903				migratetype);
1904		if (page) {
1905			preferred_zone->compact_considered = 0;
1906			preferred_zone->compact_defer_shift = 0;
1907			count_vm_event(COMPACTSUCCESS);
1908			return page;
1909		}
 
 
 
 
 
 
1910
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1911		/*
1912		 * It's bad if compaction run occurs and fails.
1913		 * The most likely reason is that pages exist,
1914		 * but not enough to satisfy watermarks.
 
 
 
 
1915		 */
1916		count_vm_event(COMPACTFAIL);
1917		defer_compaction(preferred_zone);
1918
1919		cond_resched();
 
 
 
1920	}
1921
1922	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
1923}
1924#else
1925static inline struct page *
1926__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1927	struct zonelist *zonelist, enum zone_type high_zoneidx,
1928	nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1929	int migratetype, unsigned long *did_some_progress,
1930	bool sync_migration)
1931{
 
1932	return NULL;
1933}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1934#endif /* CONFIG_COMPACTION */
1935
1936/* The really slow allocator path where we enter direct reclaim */
1937static inline struct page *
1938__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1939	struct zonelist *zonelist, enum zone_type high_zoneidx,
1940	nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1941	int migratetype, unsigned long *did_some_progress)
1942{
1943	struct page *page = NULL;
1944	struct reclaim_state reclaim_state;
1945	bool drained = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1946
1947	cond_resched();
1948
1949	/* We now go into synchronous reclaim */
1950	cpuset_memory_pressure_bump();
1951	current->flags |= PF_MEMALLOC;
1952	lockdep_set_current_reclaim_state(gfp_mask);
1953	reclaim_state.reclaimed_slab = 0;
1954	current->reclaim_state = &reclaim_state;
1955
1956	*did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1957
1958	current->reclaim_state = NULL;
1959	lockdep_clear_current_reclaim_state();
1960	current->flags &= ~PF_MEMALLOC;
1961
1962	cond_resched();
1963
1964	if (unlikely(!(*did_some_progress)))
1965		return NULL;
1966
1967	/* After successful reclaim, reconsider all zones for allocation */
1968	if (NUMA_BUILD)
1969		zlc_clear_zones_full(zonelist);
 
 
 
 
 
 
 
 
 
 
 
1970
1971retry:
1972	page = get_page_from_freelist(gfp_mask, nodemask, order,
1973					zonelist, high_zoneidx,
1974					alloc_flags, preferred_zone,
1975					migratetype);
1976
1977	/*
1978	 * If an allocation failed after direct reclaim, it could be because
1979	 * pages are pinned on the per-cpu lists. Drain them and try again
 
1980	 */
1981	if (!page && !drained) {
1982		drain_all_pages();
 
1983		drained = true;
1984		goto retry;
1985	}
 
 
1986
1987	return page;
1988}
1989
1990/*
1991 * This is called in the allocator slow-path if the allocation request is of
1992 * sufficient urgency to ignore watermarks and take other desperate measures
1993 */
1994static inline struct page *
1995__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1996	struct zonelist *zonelist, enum zone_type high_zoneidx,
1997	nodemask_t *nodemask, struct zone *preferred_zone,
1998	int migratetype)
1999{
2000	struct page *page;
2001
2002	do {
2003		page = get_page_from_freelist(gfp_mask, nodemask, order,
2004			zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2005			preferred_zone, migratetype);
2006
2007		if (!page && gfp_mask & __GFP_NOFAIL)
2008			wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2009	} while (!page && (gfp_mask & __GFP_NOFAIL));
2010
2011	return page;
2012}
2013
2014static inline
2015void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2016						enum zone_type high_zoneidx,
2017						enum zone_type classzone_idx)
2018{
2019	struct zoneref *z;
2020	struct zone *zone;
 
 
2021
2022	for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2023		wakeup_kswapd(zone, order, classzone_idx);
 
 
 
 
 
 
 
2024}
2025
2026static inline int
2027gfp_to_alloc_flags(gfp_t gfp_mask)
2028{
2029	int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2030	const gfp_t wait = gfp_mask & __GFP_WAIT;
2031
2032	/* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2033	BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
 
 
 
 
 
2034
2035	/*
2036	 * The caller may dip into page reserves a bit more if the caller
2037	 * cannot run direct reclaim, or if the caller has realtime scheduling
2038	 * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
2039	 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2040	 */
2041	alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
 
2042
2043	if (!wait) {
2044		/*
2045		 * Not worth trying to allocate harder for
2046		 * __GFP_NOMEMALLOC even if it can't schedule.
2047		 */
2048		if  (!(gfp_mask & __GFP_NOMEMALLOC))
2049			alloc_flags |= ALLOC_HARDER;
 
 
 
 
 
2050		/*
2051		 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2052		 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
 
2053		 */
2054		alloc_flags &= ~ALLOC_CPUSET;
2055	} else if (unlikely(rt_task(current)) && !in_interrupt())
2056		alloc_flags |= ALLOC_HARDER;
 
2057
2058	if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2059		if (!in_interrupt() &&
2060		    ((current->flags & PF_MEMALLOC) ||
2061		     unlikely(test_thread_flag(TIF_MEMDIE))))
2062			alloc_flags |= ALLOC_NO_WATERMARKS;
2063	}
2064
2065	return alloc_flags;
2066}
2067
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2068static inline struct page *
2069__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2070	struct zonelist *zonelist, enum zone_type high_zoneidx,
2071	nodemask_t *nodemask, struct zone *preferred_zone,
2072	int migratetype)
2073{
2074	const gfp_t wait = gfp_mask & __GFP_WAIT;
 
 
2075	struct page *page = NULL;
2076	int alloc_flags;
2077	unsigned long pages_reclaimed = 0;
2078	unsigned long did_some_progress;
2079	bool sync_migration = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
2080
2081	/*
2082	 * In the slowpath, we sanity check order to avoid ever trying to
2083	 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2084	 * be using allocators in order of preference for an area that is
2085	 * too large.
2086	 */
2087	if (order >= MAX_ORDER) {
2088		WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2089		return NULL;
2090	}
2091
2092	/*
2093	 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2094	 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2095	 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2096	 * using a larger set of nodes after it has established that the
2097	 * allowed per node queues are empty and that nodes are
2098	 * over allocated.
2099	 */
2100	if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
 
 
2101		goto nopage;
2102
2103restart:
2104	if (!(gfp_mask & __GFP_NO_KSWAPD))
2105		wake_all_kswapd(order, zonelist, high_zoneidx,
2106						zone_idx(preferred_zone));
2107
2108	/*
2109	 * OK, we're below the kswapd watermark and have kicked background
2110	 * reclaim. Now things get more complex, so set up alloc_flags according
2111	 * to how we want to proceed.
2112	 */
2113	alloc_flags = gfp_to_alloc_flags(gfp_mask);
 
 
 
 
 
 
 
 
 
2114
2115	/*
2116	 * Find the true preferred zone if the allocation is unconstrained by
2117	 * cpusets.
2118	 */
2119	if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2120		first_zones_zonelist(zonelist, high_zoneidx, NULL,
2121					&preferred_zone);
2122
2123rebalance:
2124	/* This is the last chance, in general, before the goto nopage. */
2125	page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2126			high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2127			preferred_zone, migratetype);
2128	if (page)
2129		goto got_pg;
2130
2131	/* Allocate without watermarks if the context allows */
2132	if (alloc_flags & ALLOC_NO_WATERMARKS) {
2133		page = __alloc_pages_high_priority(gfp_mask, order,
2134				zonelist, high_zoneidx, nodemask,
2135				preferred_zone, migratetype);
 
 
 
 
 
 
 
 
 
 
 
 
2136		if (page)
2137			goto got_pg;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2138	}
2139
2140	/* Atomic allocations - we can't balance anything */
2141	if (!wait)
 
 
 
 
 
2142		goto nopage;
2143
2144	/* Avoid recursion of direct reclaim */
2145	if (current->flags & PF_MEMALLOC)
2146		goto nopage;
2147
2148	/* Avoid allocations with no watermarks from looping endlessly */
2149	if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
 
 
 
 
 
 
 
 
 
 
 
 
2150		goto nopage;
2151
2152	/*
2153	 * Try direct compaction. The first pass is asynchronous. Subsequent
2154	 * attempts after direct reclaim are synchronous
2155	 */
2156	page = __alloc_pages_direct_compact(gfp_mask, order,
2157					zonelist, high_zoneidx,
2158					nodemask,
2159					alloc_flags, preferred_zone,
2160					migratetype, &did_some_progress,
2161					sync_migration);
2162	if (page)
2163		goto got_pg;
2164	sync_migration = true;
2165
2166	/* Try direct reclaim and then allocating */
2167	page = __alloc_pages_direct_reclaim(gfp_mask, order,
2168					zonelist, high_zoneidx,
2169					nodemask,
2170					alloc_flags, preferred_zone,
2171					migratetype, &did_some_progress);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2172	if (page)
2173		goto got_pg;
2174
 
 
 
 
 
 
 
 
 
 
 
 
 
2175	/*
2176	 * If we failed to make any progress reclaiming, then we are
2177	 * running out of options and have to consider going OOM
2178	 */
2179	if (!did_some_progress) {
2180		if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2181			if (oom_killer_disabled)
2182				goto nopage;
2183			page = __alloc_pages_may_oom(gfp_mask, order,
2184					zonelist, high_zoneidx,
2185					nodemask, preferred_zone,
2186					migratetype);
2187			if (page)
2188				goto got_pg;
2189
2190			if (!(gfp_mask & __GFP_NOFAIL)) {
2191				/*
2192				 * The oom killer is not called for high-order
2193				 * allocations that may fail, so if no progress
2194				 * is being made, there are no other options and
2195				 * retrying is unlikely to help.
2196				 */
2197				if (order > PAGE_ALLOC_COSTLY_ORDER)
2198					goto nopage;
2199				/*
2200				 * The oom killer is not called for lowmem
2201				 * allocations to prevent needlessly killing
2202				 * innocent tasks.
2203				 */
2204				if (high_zoneidx < ZONE_NORMAL)
2205					goto nopage;
2206			}
2207
2208			goto restart;
2209		}
2210	}
 
 
 
2211
2212	/* Check if we should retry the allocation */
2213	pages_reclaimed += did_some_progress;
2214	if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2215		/* Wait for some write requests to complete then retry */
2216		wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2217		goto rebalance;
2218	} else {
2219		/*
2220		 * High-order allocations do not necessarily loop after
2221		 * direct reclaim and reclaim/compaction depends on compaction
2222		 * being called after reclaim so call directly if necessary
 
2223		 */
2224		page = __alloc_pages_direct_compact(gfp_mask, order,
2225					zonelist, high_zoneidx,
2226					nodemask,
2227					alloc_flags, preferred_zone,
2228					migratetype, &did_some_progress,
2229					sync_migration);
 
 
 
 
2230		if (page)
2231			goto got_pg;
2232	}
2233
2234nopage:
2235	warn_alloc_failed(gfp_mask, order, NULL);
2236	return page;
 
 
 
2237got_pg:
2238	if (kmemcheck_enabled)
2239		kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2240	return page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2241
 
2242}
2243
2244/*
2245 * This is the 'heart' of the zoned buddy allocator.
2246 */
2247struct page *
2248__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2249			struct zonelist *zonelist, nodemask_t *nodemask)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2250{
2251	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2252	struct zone *preferred_zone;
2253	struct page *page;
2254	int migratetype = allocflags_to_migratetype(gfp_mask);
 
 
 
 
 
 
 
 
2255
2256	gfp_mask &= gfp_allowed_mask;
 
 
 
 
 
2257
2258	lockdep_trace_alloc(gfp_mask);
 
 
2259
2260	might_sleep_if(gfp_mask & __GFP_WAIT);
 
 
2261
2262	if (should_fail_alloc_page(gfp_mask, order))
2263		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2264
2265	/*
2266	 * Check the zones suitable for the gfp_mask contain at least one
2267	 * valid zone. It's possible to have an empty zonelist as a result
2268	 * of GFP_THISNODE and a memoryless node
2269	 */
2270	if (unlikely(!zonelist->_zonerefs->zone))
2271		return NULL;
2272
2273	get_mems_allowed();
2274	/* The preferred zone is used for statistics later */
2275	first_zones_zonelist(zonelist, high_zoneidx,
2276				nodemask ? : &cpuset_current_mems_allowed,
2277				&preferred_zone);
2278	if (!preferred_zone) {
2279		put_mems_allowed();
 
 
 
 
 
2280		return NULL;
2281	}
 
 
 
 
 
2282
2283	/* First allocation attempt */
2284	page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2285			zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2286			preferred_zone, migratetype);
2287	if (unlikely(!page))
2288		page = __alloc_pages_slowpath(gfp_mask, order,
2289				zonelist, high_zoneidx, nodemask,
2290				preferred_zone, migratetype);
2291	put_mems_allowed();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2292
2293	trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2294	return page;
2295}
2296EXPORT_SYMBOL(__alloc_pages_nodemask);
 
 
 
 
 
 
 
 
 
2297
2298/*
2299 * Common helper functions.
 
 
2300 */
2301unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2302{
2303	struct page *page;
2304
2305	/*
2306	 * __get_free_pages() returns a 32-bit address, which cannot represent
2307	 * a highmem page
2308	 */
2309	VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2310
2311	page = alloc_pages(gfp_mask, order);
2312	if (!page)
2313		return 0;
2314	return (unsigned long) page_address(page);
2315}
2316EXPORT_SYMBOL(__get_free_pages);
2317
2318unsigned long get_zeroed_page(gfp_t gfp_mask)
2319{
2320	return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2321}
2322EXPORT_SYMBOL(get_zeroed_page);
2323
2324void __pagevec_free(struct pagevec *pvec)
2325{
2326	int i = pagevec_count(pvec);
2327
2328	while (--i >= 0) {
2329		trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2330		free_hot_cold_page(pvec->pages[i], pvec->cold);
2331	}
2332}
2333
 
 
 
 
 
 
 
 
 
 
2334void __free_pages(struct page *page, unsigned int order)
2335{
2336	if (put_page_testzero(page)) {
2337		if (order == 0)
2338			free_hot_cold_page(page, 0);
2339		else
2340			__free_pages_ok(page, order);
2341	}
2342}
2343
 
 
 
 
 
 
2344EXPORT_SYMBOL(__free_pages);
2345
2346void free_pages(unsigned long addr, unsigned int order)
2347{
2348	if (addr != 0) {
2349		VM_BUG_ON(!virt_addr_valid((void *)addr));
2350		__free_pages(virt_to_page((void *)addr), order);
2351	}
2352}
2353
2354EXPORT_SYMBOL(free_pages);
2355
2356static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
 
 
 
 
 
 
 
 
 
 
 
 
2357{
2358	if (addr) {
2359		unsigned long alloc_end = addr + (PAGE_SIZE << order);
2360		unsigned long used = addr + PAGE_ALIGN(size);
2361
2362		split_page(virt_to_page((void *)addr), order);
2363		while (used < alloc_end) {
2364			free_page(used);
2365			used += PAGE_SIZE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2366		}
2367	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2368	return (void *)addr;
2369}
2370
2371/**
2372 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2373 * @size: the number of bytes to allocate
2374 * @gfp_mask: GFP flags for the allocation
2375 *
2376 * This function is similar to alloc_pages(), except that it allocates the
2377 * minimum number of pages to satisfy the request.  alloc_pages() can only
2378 * allocate memory in power-of-two pages.
2379 *
2380 * This function is also limited by MAX_ORDER.
2381 *
2382 * Memory allocated by this function must be released by free_pages_exact().
 
 
2383 */
2384void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2385{
2386	unsigned int order = get_order(size);
2387	unsigned long addr;
2388
 
 
 
2389	addr = __get_free_pages(gfp_mask, order);
2390	return make_alloc_exact(addr, order, size);
2391}
2392EXPORT_SYMBOL(alloc_pages_exact);
2393
2394/**
2395 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2396 *			   pages on a node.
2397 * @nid: the preferred node ID where memory should be allocated
2398 * @size: the number of bytes to allocate
2399 * @gfp_mask: GFP flags for the allocation
2400 *
2401 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2402 * back.
2403 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2404 * but is not exact.
2405 */
2406void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2407{
2408	unsigned order = get_order(size);
2409	struct page *p = alloc_pages_node(nid, gfp_mask, order);
 
 
 
 
 
2410	if (!p)
2411		return NULL;
2412	return make_alloc_exact((unsigned long)page_address(p), order, size);
2413}
2414EXPORT_SYMBOL(alloc_pages_exact_nid);
2415
2416/**
2417 * free_pages_exact - release memory allocated via alloc_pages_exact()
2418 * @virt: the value returned by alloc_pages_exact.
2419 * @size: size of allocation, same value as passed to alloc_pages_exact().
2420 *
2421 * Release the memory allocated by a previous call to alloc_pages_exact.
2422 */
2423void free_pages_exact(void *virt, size_t size)
2424{
2425	unsigned long addr = (unsigned long)virt;
2426	unsigned long end = addr + PAGE_ALIGN(size);
2427
2428	while (addr < end) {
2429		free_page(addr);
2430		addr += PAGE_SIZE;
2431	}
2432}
2433EXPORT_SYMBOL(free_pages_exact);
2434
2435static unsigned int nr_free_zone_pages(int offset)
 
 
 
 
 
 
 
 
 
 
 
 
2436{
2437	struct zoneref *z;
2438	struct zone *zone;
2439
2440	/* Just pick one node, since fallback list is circular */
2441	unsigned int sum = 0;
2442
2443	struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2444
2445	for_each_zone_zonelist(zone, z, zonelist, offset) {
2446		unsigned long size = zone->present_pages;
2447		unsigned long high = high_wmark_pages(zone);
2448		if (size > high)
2449			sum += size - high;
2450	}
2451
2452	return sum;
2453}
2454
2455/*
2456 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
 
 
 
 
 
 
2457 */
2458unsigned int nr_free_buffer_pages(void)
2459{
2460	return nr_free_zone_pages(gfp_zone(GFP_USER));
2461}
2462EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2463
2464/*
2465 * Amount of free RAM allocatable within all zones
2466 */
2467unsigned int nr_free_pagecache_pages(void)
2468{
2469	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2470}
2471
2472static inline void show_node(struct zone *zone)
2473{
2474	if (NUMA_BUILD)
2475		printk("Node %d ", zone_to_nid(zone));
2476}
2477
2478void si_meminfo(struct sysinfo *val)
2479{
2480	val->totalram = totalram_pages;
2481	val->sharedram = 0;
2482	val->freeram = global_page_state(NR_FREE_PAGES);
2483	val->bufferram = nr_blockdev_pages();
2484	val->totalhigh = totalhigh_pages;
2485	val->freehigh = nr_free_highpages();
2486	val->mem_unit = PAGE_SIZE;
2487}
2488
2489EXPORT_SYMBOL(si_meminfo);
2490
2491#ifdef CONFIG_NUMA
2492void si_meminfo_node(struct sysinfo *val, int nid)
2493{
2494	pg_data_t *pgdat = NODE_DATA(nid);
2495
2496	val->totalram = pgdat->node_present_pages;
2497	val->freeram = node_page_state(nid, NR_FREE_PAGES);
2498#ifdef CONFIG_HIGHMEM
2499	val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2500	val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2501			NR_FREE_PAGES);
2502#else
2503	val->totalhigh = 0;
2504	val->freehigh = 0;
2505#endif
2506	val->mem_unit = PAGE_SIZE;
2507}
2508#endif
2509
2510/*
2511 * Determine whether the node should be displayed or not, depending on whether
2512 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2513 */
2514bool skip_free_areas_node(unsigned int flags, int nid)
2515{
2516	bool ret = false;
2517
2518	if (!(flags & SHOW_MEM_FILTER_NODES))
2519		goto out;
2520
2521	get_mems_allowed();
2522	ret = !node_isset(nid, cpuset_current_mems_allowed);
2523	put_mems_allowed();
2524out:
2525	return ret;
2526}
2527
2528#define K(x) ((x) << (PAGE_SHIFT-10))
2529
2530/*
2531 * Show free area list (used inside shift_scroll-lock stuff)
2532 * We also calculate the percentage fragmentation. We do this by counting the
2533 * memory on each free list with the exception of the first item on the list.
2534 * Suppresses nodes that are not allowed by current's cpuset if
2535 * SHOW_MEM_FILTER_NODES is passed.
2536 */
2537void show_free_areas(unsigned int filter)
2538{
2539	int cpu;
2540	struct zone *zone;
2541
2542	for_each_populated_zone(zone) {
2543		if (skip_free_areas_node(filter, zone_to_nid(zone)))
2544			continue;
2545		show_node(zone);
2546		printk("%s per-cpu:\n", zone->name);
2547
2548		for_each_online_cpu(cpu) {
2549			struct per_cpu_pageset *pageset;
2550
2551			pageset = per_cpu_ptr(zone->pageset, cpu);
2552
2553			printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2554			       cpu, pageset->pcp.high,
2555			       pageset->pcp.batch, pageset->pcp.count);
2556		}
2557	}
2558
2559	printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2560		" active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2561		" unevictable:%lu"
2562		" dirty:%lu writeback:%lu unstable:%lu\n"
2563		" free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2564		" mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2565		global_page_state(NR_ACTIVE_ANON),
2566		global_page_state(NR_INACTIVE_ANON),
2567		global_page_state(NR_ISOLATED_ANON),
2568		global_page_state(NR_ACTIVE_FILE),
2569		global_page_state(NR_INACTIVE_FILE),
2570		global_page_state(NR_ISOLATED_FILE),
2571		global_page_state(NR_UNEVICTABLE),
2572		global_page_state(NR_FILE_DIRTY),
2573		global_page_state(NR_WRITEBACK),
2574		global_page_state(NR_UNSTABLE_NFS),
2575		global_page_state(NR_FREE_PAGES),
2576		global_page_state(NR_SLAB_RECLAIMABLE),
2577		global_page_state(NR_SLAB_UNRECLAIMABLE),
2578		global_page_state(NR_FILE_MAPPED),
2579		global_page_state(NR_SHMEM),
2580		global_page_state(NR_PAGETABLE),
2581		global_page_state(NR_BOUNCE));
2582
2583	for_each_populated_zone(zone) {
2584		int i;
2585
2586		if (skip_free_areas_node(filter, zone_to_nid(zone)))
2587			continue;
2588		show_node(zone);
2589		printk("%s"
2590			" free:%lukB"
2591			" min:%lukB"
2592			" low:%lukB"
2593			" high:%lukB"
2594			" active_anon:%lukB"
2595			" inactive_anon:%lukB"
2596			" active_file:%lukB"
2597			" inactive_file:%lukB"
2598			" unevictable:%lukB"
2599			" isolated(anon):%lukB"
2600			" isolated(file):%lukB"
2601			" present:%lukB"
2602			" mlocked:%lukB"
2603			" dirty:%lukB"
2604			" writeback:%lukB"
2605			" mapped:%lukB"
2606			" shmem:%lukB"
2607			" slab_reclaimable:%lukB"
2608			" slab_unreclaimable:%lukB"
2609			" kernel_stack:%lukB"
2610			" pagetables:%lukB"
2611			" unstable:%lukB"
2612			" bounce:%lukB"
2613			" writeback_tmp:%lukB"
2614			" pages_scanned:%lu"
2615			" all_unreclaimable? %s"
2616			"\n",
2617			zone->name,
2618			K(zone_page_state(zone, NR_FREE_PAGES)),
2619			K(min_wmark_pages(zone)),
2620			K(low_wmark_pages(zone)),
2621			K(high_wmark_pages(zone)),
2622			K(zone_page_state(zone, NR_ACTIVE_ANON)),
2623			K(zone_page_state(zone, NR_INACTIVE_ANON)),
2624			K(zone_page_state(zone, NR_ACTIVE_FILE)),
2625			K(zone_page_state(zone, NR_INACTIVE_FILE)),
2626			K(zone_page_state(zone, NR_UNEVICTABLE)),
2627			K(zone_page_state(zone, NR_ISOLATED_ANON)),
2628			K(zone_page_state(zone, NR_ISOLATED_FILE)),
2629			K(zone->present_pages),
2630			K(zone_page_state(zone, NR_MLOCK)),
2631			K(zone_page_state(zone, NR_FILE_DIRTY)),
2632			K(zone_page_state(zone, NR_WRITEBACK)),
2633			K(zone_page_state(zone, NR_FILE_MAPPED)),
2634			K(zone_page_state(zone, NR_SHMEM)),
2635			K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2636			K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2637			zone_page_state(zone, NR_KERNEL_STACK) *
2638				THREAD_SIZE / 1024,
2639			K(zone_page_state(zone, NR_PAGETABLE)),
2640			K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2641			K(zone_page_state(zone, NR_BOUNCE)),
2642			K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2643			zone->pages_scanned,
2644			(zone->all_unreclaimable ? "yes" : "no")
2645			);
2646		printk("lowmem_reserve[]:");
2647		for (i = 0; i < MAX_NR_ZONES; i++)
2648			printk(" %lu", zone->lowmem_reserve[i]);
2649		printk("\n");
2650	}
2651
2652	for_each_populated_zone(zone) {
2653 		unsigned long nr[MAX_ORDER], flags, order, total = 0;
2654
2655		if (skip_free_areas_node(filter, zone_to_nid(zone)))
2656			continue;
2657		show_node(zone);
2658		printk("%s: ", zone->name);
2659
2660		spin_lock_irqsave(&zone->lock, flags);
2661		for (order = 0; order < MAX_ORDER; order++) {
2662			nr[order] = zone->free_area[order].nr_free;
2663			total += nr[order] << order;
2664		}
2665		spin_unlock_irqrestore(&zone->lock, flags);
2666		for (order = 0; order < MAX_ORDER; order++)
2667			printk("%lu*%lukB ", nr[order], K(1UL) << order);
2668		printk("= %lukB\n", K(total));
2669	}
2670
2671	printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2672
2673	show_swap_cache_info();
2674}
2675
2676static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2677{
2678	zoneref->zone = zone;
2679	zoneref->zone_idx = zone_idx(zone);
2680}
2681
2682/*
2683 * Builds allocation fallback zone lists.
2684 *
2685 * Add all populated zones of a node to the zonelist.
2686 */
2687static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2688				int nr_zones, enum zone_type zone_type)
2689{
2690	struct zone *zone;
2691
2692	BUG_ON(zone_type >= MAX_NR_ZONES);
2693	zone_type++;
2694
2695	do {
2696		zone_type--;
2697		zone = pgdat->node_zones + zone_type;
2698		if (populated_zone(zone)) {
2699			zoneref_set_zone(zone,
2700				&zonelist->_zonerefs[nr_zones++]);
2701			check_highest_zone(zone_type);
2702		}
2703
2704	} while (zone_type);
 
2705	return nr_zones;
2706}
2707
2708
2709/*
2710 *  zonelist_order:
2711 *  0 = automatic detection of better ordering.
2712 *  1 = order by ([node] distance, -zonetype)
2713 *  2 = order by (-zonetype, [node] distance)
2714 *
2715 *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2716 *  the same zonelist. So only NUMA can configure this param.
2717 */
2718#define ZONELIST_ORDER_DEFAULT  0
2719#define ZONELIST_ORDER_NODE     1
2720#define ZONELIST_ORDER_ZONE     2
2721
2722/* zonelist order in the kernel.
2723 * set_zonelist_order() will set this to NODE or ZONE.
2724 */
2725static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2726static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2727
2728
2729#ifdef CONFIG_NUMA
2730/* The value user specified ....changed by config */
2731static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2732/* string for sysctl */
2733#define NUMA_ZONELIST_ORDER_LEN	16
2734char numa_zonelist_order[16] = "default";
2735
2736/*
2737 * interface for configure zonelist ordering.
2738 * command line option "numa_zonelist_order"
2739 *	= "[dD]efault	- default, automatic configuration.
2740 *	= "[nN]ode 	- order by node locality, then by zone within node
2741 *	= "[zZ]one      - order by zone, then by locality within zone
2742 */
2743
2744static int __parse_numa_zonelist_order(char *s)
2745{
2746	if (*s == 'd' || *s == 'D') {
2747		user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2748	} else if (*s == 'n' || *s == 'N') {
2749		user_zonelist_order = ZONELIST_ORDER_NODE;
2750	} else if (*s == 'z' || *s == 'Z') {
2751		user_zonelist_order = ZONELIST_ORDER_ZONE;
2752	} else {
2753		printk(KERN_WARNING
2754			"Ignoring invalid numa_zonelist_order value:  "
2755			"%s\n", s);
2756		return -EINVAL;
2757	}
2758	return 0;
2759}
2760
2761static __init int setup_numa_zonelist_order(char *s)
2762{
2763	int ret;
2764
2765	if (!s)
2766		return 0;
2767
2768	ret = __parse_numa_zonelist_order(s);
2769	if (ret == 0)
2770		strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2771
2772	return ret;
2773}
2774early_param("numa_zonelist_order", setup_numa_zonelist_order);
2775
2776/*
2777 * sysctl handler for numa_zonelist_order
2778 */
2779int numa_zonelist_order_handler(ctl_table *table, int write,
2780		void __user *buffer, size_t *length,
2781		loff_t *ppos)
2782{
2783	char saved_string[NUMA_ZONELIST_ORDER_LEN];
2784	int ret;
2785	static DEFINE_MUTEX(zl_order_mutex);
2786
2787	mutex_lock(&zl_order_mutex);
2788	if (write)
2789		strcpy(saved_string, (char*)table->data);
2790	ret = proc_dostring(table, write, buffer, length, ppos);
2791	if (ret)
2792		goto out;
2793	if (write) {
2794		int oldval = user_zonelist_order;
2795		if (__parse_numa_zonelist_order((char*)table->data)) {
2796			/*
2797			 * bogus value.  restore saved string
2798			 */
2799			strncpy((char*)table->data, saved_string,
2800				NUMA_ZONELIST_ORDER_LEN);
2801			user_zonelist_order = oldval;
2802		} else if (oldval != user_zonelist_order) {
2803			mutex_lock(&zonelists_mutex);
2804			build_all_zonelists(NULL);
2805			mutex_unlock(&zonelists_mutex);
2806		}
2807	}
2808out:
2809	mutex_unlock(&zl_order_mutex);
2810	return ret;
2811}
2812
2813
2814#define MAX_NODE_LOAD (nr_online_nodes)
2815static int node_load[MAX_NUMNODES];
2816
2817/**
2818 * find_next_best_node - find the next node that should appear in a given node's fallback list
2819 * @node: node whose fallback list we're appending
2820 * @used_node_mask: nodemask_t of already used nodes
2821 *
2822 * We use a number of factors to determine which is the next node that should
2823 * appear on a given node's fallback list.  The node should not have appeared
2824 * already in @node's fallback list, and it should be the next closest node
2825 * according to the distance array (which contains arbitrary distance values
2826 * from each node to each node in the system), and should also prefer nodes
2827 * with no CPUs, since presumably they'll have very little allocation pressure
2828 * on them otherwise.
2829 * It returns -1 if no node is found.
 
2830 */
2831static int find_next_best_node(int node, nodemask_t *used_node_mask)
2832{
2833	int n, val;
2834	int min_val = INT_MAX;
2835	int best_node = -1;
2836	const struct cpumask *tmp = cpumask_of_node(0);
2837
2838	/* Use the local node if we haven't already */
2839	if (!node_isset(node, *used_node_mask)) {
 
 
 
2840		node_set(node, *used_node_mask);
2841		return node;
2842	}
2843
2844	for_each_node_state(n, N_HIGH_MEMORY) {
2845
2846		/* Don't want a node to appear more than once */
2847		if (node_isset(n, *used_node_mask))
2848			continue;
2849
2850		/* Use the distance array to find the distance */
2851		val = node_distance(node, n);
2852
2853		/* Penalize nodes under us ("prefer the next node") */
2854		val += (n < node);
2855
2856		/* Give preference to headless and unused nodes */
2857		tmp = cpumask_of_node(n);
2858		if (!cpumask_empty(tmp))
2859			val += PENALTY_FOR_NODE_WITH_CPUS;
2860
2861		/* Slight preference for less loaded node */
2862		val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2863		val += node_load[n];
2864
2865		if (val < min_val) {
2866			min_val = val;
2867			best_node = n;
2868		}
2869	}
2870
2871	if (best_node >= 0)
2872		node_set(best_node, *used_node_mask);
2873
2874	return best_node;
2875}
2876
2877
2878/*
2879 * Build zonelists ordered by node and zones within node.
2880 * This results in maximum locality--normal zone overflows into local
2881 * DMA zone, if any--but risks exhausting DMA zone.
2882 */
2883static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
 
2884{
2885	int j;
2886	struct zonelist *zonelist;
 
 
 
 
 
 
 
2887
2888	zonelist = &pgdat->node_zonelists[0];
2889	for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2890		;
2891	j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2892							MAX_NR_ZONES - 1);
2893	zonelist->_zonerefs[j].zone = NULL;
2894	zonelist->_zonerefs[j].zone_idx = 0;
2895}
2896
2897/*
2898 * Build gfp_thisnode zonelists
2899 */
2900static void build_thisnode_zonelists(pg_data_t *pgdat)
2901{
2902	int j;
2903	struct zonelist *zonelist;
2904
2905	zonelist = &pgdat->node_zonelists[1];
2906	j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2907	zonelist->_zonerefs[j].zone = NULL;
2908	zonelist->_zonerefs[j].zone_idx = 0;
 
2909}
2910
2911/*
2912 * Build zonelists ordered by zone and nodes within zones.
2913 * This results in conserving DMA zone[s] until all Normal memory is
2914 * exhausted, but results in overflowing to remote node while memory
2915 * may still exist in local DMA zone.
2916 */
2917static int node_order[MAX_NUMNODES];
2918
2919static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2920{
2921	int pos, j, node;
2922	int zone_type;		/* needs to be signed */
2923	struct zone *z;
2924	struct zonelist *zonelist;
2925
2926	zonelist = &pgdat->node_zonelists[0];
2927	pos = 0;
2928	for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2929		for (j = 0; j < nr_nodes; j++) {
2930			node = node_order[j];
2931			z = &NODE_DATA(node)->node_zones[zone_type];
2932			if (populated_zone(z)) {
2933				zoneref_set_zone(z,
2934					&zonelist->_zonerefs[pos++]);
2935				check_highest_zone(zone_type);
2936			}
2937		}
2938	}
2939	zonelist->_zonerefs[pos].zone = NULL;
2940	zonelist->_zonerefs[pos].zone_idx = 0;
2941}
2942
2943static int default_zonelist_order(void)
2944{
2945	int nid, zone_type;
2946	unsigned long low_kmem_size,total_size;
2947	struct zone *z;
2948	int average_size;
2949	/*
2950         * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2951	 * If they are really small and used heavily, the system can fall
2952	 * into OOM very easily.
2953	 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2954	 */
2955	/* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2956	low_kmem_size = 0;
2957	total_size = 0;
2958	for_each_online_node(nid) {
2959		for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2960			z = &NODE_DATA(nid)->node_zones[zone_type];
2961			if (populated_zone(z)) {
2962				if (zone_type < ZONE_NORMAL)
2963					low_kmem_size += z->present_pages;
2964				total_size += z->present_pages;
2965			} else if (zone_type == ZONE_NORMAL) {
2966				/*
2967				 * If any node has only lowmem, then node order
2968				 * is preferred to allow kernel allocations
2969				 * locally; otherwise, they can easily infringe
2970				 * on other nodes when there is an abundance of
2971				 * lowmem available to allocate from.
2972				 */
2973				return ZONELIST_ORDER_NODE;
2974			}
2975		}
2976	}
2977	if (!low_kmem_size ||  /* there are no DMA area. */
2978	    low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2979		return ZONELIST_ORDER_NODE;
2980	/*
2981	 * look into each node's config.
2982  	 * If there is a node whose DMA/DMA32 memory is very big area on
2983 	 * local memory, NODE_ORDER may be suitable.
2984         */
2985	average_size = total_size /
2986				(nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2987	for_each_online_node(nid) {
2988		low_kmem_size = 0;
2989		total_size = 0;
2990		for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2991			z = &NODE_DATA(nid)->node_zones[zone_type];
2992			if (populated_zone(z)) {
2993				if (zone_type < ZONE_NORMAL)
2994					low_kmem_size += z->present_pages;
2995				total_size += z->present_pages;
2996			}
2997		}
2998		if (low_kmem_size &&
2999		    total_size > average_size && /* ignore small node */
3000		    low_kmem_size > total_size * 70/100)
3001			return ZONELIST_ORDER_NODE;
3002	}
3003	return ZONELIST_ORDER_ZONE;
3004}
3005
3006static void set_zonelist_order(void)
3007{
3008	if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3009		current_zonelist_order = default_zonelist_order();
3010	else
3011		current_zonelist_order = user_zonelist_order;
3012}
3013
3014static void build_zonelists(pg_data_t *pgdat)
3015{
3016	int j, node, load;
3017	enum zone_type i;
3018	nodemask_t used_mask;
3019	int local_node, prev_node;
3020	struct zonelist *zonelist;
3021	int order = current_zonelist_order;
3022
3023	/* initialize zonelists */
3024	for (i = 0; i < MAX_ZONELISTS; i++) {
3025		zonelist = pgdat->node_zonelists + i;
3026		zonelist->_zonerefs[0].zone = NULL;
3027		zonelist->_zonerefs[0].zone_idx = 0;
3028	}
3029
3030	/* NUMA-aware ordering of nodes */
3031	local_node = pgdat->node_id;
3032	load = nr_online_nodes;
3033	prev_node = local_node;
3034	nodes_clear(used_mask);
3035
3036	memset(node_order, 0, sizeof(node_order));
3037	j = 0;
3038
3039	while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3040		int distance = node_distance(local_node, node);
3041
3042		/*
3043		 * If another node is sufficiently far away then it is better
3044		 * to reclaim pages in a zone before going off node.
3045		 */
3046		if (distance > RECLAIM_DISTANCE)
3047			zone_reclaim_mode = 1;
3048
3049		/*
3050		 * We don't want to pressure a particular node.
3051		 * So adding penalty to the first node in same
3052		 * distance group to make it round-robin.
3053		 */
3054		if (distance != node_distance(local_node, prev_node))
3055			node_load[node] = load;
 
3056
 
3057		prev_node = node;
3058		load--;
3059		if (order == ZONELIST_ORDER_NODE)
3060			build_zonelists_in_node_order(pgdat, node);
3061		else
3062			node_order[j++] = node;	/* remember order */
3063	}
3064
3065	if (order == ZONELIST_ORDER_ZONE) {
3066		/* calculate node order -- i.e., DMA last! */
3067		build_zonelists_in_zone_order(pgdat, j);
3068	}
3069
 
3070	build_thisnode_zonelists(pgdat);
3071}
3072
3073/* Construct the zonelist performance cache - see further mmzone.h */
3074static void build_zonelist_cache(pg_data_t *pgdat)
3075{
3076	struct zonelist *zonelist;
3077	struct zonelist_cache *zlc;
3078	struct zoneref *z;
3079
3080	zonelist = &pgdat->node_zonelists[0];
3081	zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3082	bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3083	for (z = zonelist->_zonerefs; z->zone; z++)
3084		zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3085}
3086
3087#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3088/*
3089 * Return node id of node used for "local" allocations.
3090 * I.e., first node id of first zone in arg node's generic zonelist.
3091 * Used for initializing percpu 'numa_mem', which is used primarily
3092 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3093 */
3094int local_memory_node(int node)
3095{
3096	struct zone *zone;
3097
3098	(void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3099				   gfp_zone(GFP_KERNEL),
3100				   NULL,
3101				   &zone);
3102	return zone->node;
3103}
3104#endif
3105
 
 
3106#else	/* CONFIG_NUMA */
3107
3108static void set_zonelist_order(void)
3109{
3110	current_zonelist_order = ZONELIST_ORDER_ZONE;
3111}
3112
3113static void build_zonelists(pg_data_t *pgdat)
3114{
3115	int node, local_node;
3116	enum zone_type j;
3117	struct zonelist *zonelist;
3118
3119	local_node = pgdat->node_id;
3120
3121	zonelist = &pgdat->node_zonelists[0];
3122	j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
 
3123
3124	/*
3125	 * Now we build the zonelist so that it contains the zones
3126	 * of all the other nodes.
3127	 * We don't want to pressure a particular node, so when
3128	 * building the zones for node N, we make sure that the
3129	 * zones coming right after the local ones are those from
3130	 * node N+1 (modulo N)
3131	 */
3132	for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3133		if (!node_online(node))
3134			continue;
3135		j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3136							MAX_NR_ZONES - 1);
3137	}
3138	for (node = 0; node < local_node; node++) {
3139		if (!node_online(node))
3140			continue;
3141		j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3142							MAX_NR_ZONES - 1);
3143	}
3144
3145	zonelist->_zonerefs[j].zone = NULL;
3146	zonelist->_zonerefs[j].zone_idx = 0;
3147}
3148
3149/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3150static void build_zonelist_cache(pg_data_t *pgdat)
3151{
3152	pgdat->node_zonelists[0].zlcache_ptr = NULL;
3153}
3154
3155#endif	/* CONFIG_NUMA */
3156
3157/*
3158 * Boot pageset table. One per cpu which is going to be used for all
3159 * zones and all nodes. The parameters will be set in such a way
3160 * that an item put on a list will immediately be handed over to
3161 * the buddy list. This is safe since pageset manipulation is done
3162 * with interrupts disabled.
3163 *
3164 * The boot_pagesets must be kept even after bootup is complete for
3165 * unused processors and/or zones. They do play a role for bootstrapping
3166 * hotplugged processors.
3167 *
3168 * zoneinfo_show() and maybe other functions do
3169 * not check if the processor is online before following the pageset pointer.
3170 * Other parts of the kernel may not check if the zone is available.
3171 */
3172static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3173static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3174static void setup_zone_pageset(struct zone *zone);
 
 
 
3175
3176/*
3177 * Global mutex to protect against size modification of zonelists
3178 * as well as to serialize pageset setup for the new populated zone.
3179 */
3180DEFINE_MUTEX(zonelists_mutex);
3181
3182/* return values int ....just for stop_machine() */
3183static __init_refok int __build_all_zonelists(void *data)
3184{
3185	int nid;
3186	int cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3187
3188#ifdef CONFIG_NUMA
3189	memset(node_load, 0, sizeof(node_load));
3190#endif
3191	for_each_online_node(nid) {
3192		pg_data_t *pgdat = NODE_DATA(nid);
3193
3194		build_zonelists(pgdat);
3195		build_zonelist_cache(pgdat);
3196	}
3197
3198	/*
3199	 * Initialize the boot_pagesets that are going to be used
3200	 * for bootstrapping processors. The real pagesets for
3201	 * each zone will be allocated later when the per cpu
3202	 * allocator is available.
3203	 *
3204	 * boot_pagesets are used also for bootstrapping offline
3205	 * cpus if the system is already booted because the pagesets
3206	 * are needed to initialize allocators on a specific cpu too.
3207	 * F.e. the percpu allocator needs the page allocator which
3208	 * needs the percpu allocator in order to allocate its pagesets
3209	 * (a chicken-egg dilemma).
3210	 */
3211	for_each_possible_cpu(cpu) {
3212		setup_pageset(&per_cpu(boot_pageset, cpu), 0);
 
 
 
 
 
 
 
 
 
 
3213
3214#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3215		/*
3216		 * We now know the "local memory node" for each node--
3217		 * i.e., the node of the first zone in the generic zonelist.
3218		 * Set up numa_mem percpu variable for on-line cpus.  During
3219		 * boot, only the boot cpu should be on-line;  we'll init the
3220		 * secondary cpus' numa_mem as they come on-line.  During
3221		 * node/memory hotplug, we'll fixup all on-line cpus.
3222		 */
3223		if (cpu_online(cpu))
3224			set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3225#endif
3226	}
3227
3228	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3229}
3230
3231/*
3232 * Called with zonelists_mutex held always
3233 * unless system_state == SYSTEM_BOOTING.
 
 
 
3234 */
3235void __ref build_all_zonelists(void *data)
3236{
3237	set_zonelist_order();
3238
3239	if (system_state == SYSTEM_BOOTING) {
3240		__build_all_zonelists(NULL);
3241		mminit_verify_zonelist();
3242		cpuset_init_current_mems_allowed();
3243	} else {
3244		/* we have to stop all cpus to guarantee there is no user
3245		   of zonelist */
3246#ifdef CONFIG_MEMORY_HOTPLUG
3247		if (data)
3248			setup_zone_pageset((struct zone *)data);
3249#endif
3250		stop_machine(__build_all_zonelists, NULL, NULL);
3251		/* cpuset refresh routine should be here */
3252	}
3253	vm_total_pages = nr_free_pagecache_pages();
 
3254	/*
3255	 * Disable grouping by mobility if the number of pages in the
3256	 * system is too low to allow the mechanism to work. It would be
3257	 * more accurate, but expensive to check per-zone. This check is
3258	 * made on memory-hotadd so a system can start with mobility
3259	 * disabled and enable it later
3260	 */
3261	if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3262		page_group_by_mobility_disabled = 1;
3263	else
3264		page_group_by_mobility_disabled = 0;
3265
3266	printk("Built %i zonelists in %s order, mobility grouping %s.  "
3267		"Total pages: %ld\n",
3268			nr_online_nodes,
3269			zonelist_order_name[current_zonelist_order],
3270			page_group_by_mobility_disabled ? "off" : "on",
3271			vm_total_pages);
3272#ifdef CONFIG_NUMA
3273	printk("Policy zone: %s\n", zone_names[policy_zone]);
3274#endif
3275}
3276
3277/*
3278 * Helper functions to size the waitqueue hash table.
3279 * Essentially these want to choose hash table sizes sufficiently
3280 * large so that collisions trying to wait on pages are rare.
3281 * But in fact, the number of active page waitqueues on typical
3282 * systems is ridiculously low, less than 200. So this is even
3283 * conservative, even though it seems large.
3284 *
3285 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3286 * waitqueues, i.e. the size of the waitq table given the number of pages.
3287 */
3288#define PAGES_PER_WAITQUEUE	256
3289
3290#ifndef CONFIG_MEMORY_HOTPLUG
3291static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3292{
3293	unsigned long size = 1;
3294
3295	pages /= PAGES_PER_WAITQUEUE;
3296
3297	while (size < pages)
3298		size <<= 1;
3299
3300	/*
3301	 * Once we have dozens or even hundreds of threads sleeping
3302	 * on IO we've got bigger problems than wait queue collision.
3303	 * Limit the size of the wait table to a reasonable size.
3304	 */
3305	size = min(size, 4096UL);
3306
3307	return max(size, 4UL);
3308}
3309#else
3310/*
3311 * A zone's size might be changed by hot-add, so it is not possible to determine
3312 * a suitable size for its wait_table.  So we use the maximum size now.
3313 *
3314 * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
3315 *
3316 *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
3317 *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3318 *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
3319 *
3320 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3321 * or more by the traditional way. (See above).  It equals:
3322 *
3323 *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
3324 *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
3325 *    powerpc (64K page size)             : =  (32G +16M)byte.
3326 */
3327static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3328{
3329	return 4096UL;
3330}
3331#endif
3332
3333/*
3334 * This is an integer logarithm so that shifts can be used later
3335 * to extract the more random high bits from the multiplicative
3336 * hash function before the remainder is taken.
3337 */
3338static inline unsigned long wait_table_bits(unsigned long size)
3339{
3340	return ffz(~size);
3341}
3342
3343#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3344
3345/*
3346 * Check if a pageblock contains reserved pages
3347 */
3348static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3349{
3350	unsigned long pfn;
3351
3352	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3353		if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3354			return 1;
3355	}
3356	return 0;
3357}
3358
3359/*
3360 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3361 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3362 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3363 * higher will lead to a bigger reserve which will get freed as contiguous
3364 * blocks as reclaim kicks in
3365 */
3366static void setup_zone_migrate_reserve(struct zone *zone)
3367{
3368	unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3369	struct page *page;
3370	unsigned long block_migratetype;
3371	int reserve;
3372
3373	/* Get the start pfn, end pfn and the number of blocks to reserve */
3374	start_pfn = zone->zone_start_pfn;
3375	end_pfn = start_pfn + zone->spanned_pages;
3376	reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3377							pageblock_order;
3378
3379	/*
3380	 * Reserve blocks are generally in place to help high-order atomic
3381	 * allocations that are short-lived. A min_free_kbytes value that
3382	 * would result in more than 2 reserve blocks for atomic allocations
3383	 * is assumed to be in place to help anti-fragmentation for the
3384	 * future allocation of hugepages at runtime.
3385	 */
3386	reserve = min(2, reserve);
3387
3388	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3389		if (!pfn_valid(pfn))
3390			continue;
3391		page = pfn_to_page(pfn);
3392
3393		/* Watch out for overlapping nodes */
3394		if (page_to_nid(page) != zone_to_nid(zone))
3395			continue;
3396
3397		/* Blocks with reserved pages will never free, skip them. */
3398		block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3399		if (pageblock_is_reserved(pfn, block_end_pfn))
3400			continue;
3401
3402		block_migratetype = get_pageblock_migratetype(page);
3403
3404		/* If this block is reserved, account for it */
3405		if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3406			reserve--;
3407			continue;
3408		}
3409
3410		/* Suitable for reserving if this block is movable */
3411		if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3412			set_pageblock_migratetype(page, MIGRATE_RESERVE);
3413			move_freepages_block(zone, page, MIGRATE_RESERVE);
3414			reserve--;
3415			continue;
3416		}
3417
3418		/*
3419		 * If the reserve is met and this is a previous reserved block,
3420		 * take it back
3421		 */
3422		if (block_migratetype == MIGRATE_RESERVE) {
3423			set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3424			move_freepages_block(zone, page, MIGRATE_MOVABLE);
3425		}
3426	}
3427}
3428
3429/*
3430 * Initially all pages are reserved - free ones are freed
3431 * up by free_all_bootmem() once the early boot process is
3432 * done. Non-atomic initialization, single-pass.
3433 */
3434void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3435		unsigned long start_pfn, enum memmap_context context)
3436{
3437	struct page *page;
3438	unsigned long end_pfn = start_pfn + size;
3439	unsigned long pfn;
3440	struct zone *z;
3441
3442	if (highest_memmap_pfn < end_pfn - 1)
3443		highest_memmap_pfn = end_pfn - 1;
3444
3445	z = &NODE_DATA(nid)->node_zones[zone];
3446	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3447		/*
3448		 * There can be holes in boot-time mem_map[]s
3449		 * handed to this function.  They do not
3450		 * exist on hotplugged memory.
3451		 */
3452		if (context == MEMMAP_EARLY) {
3453			if (!early_pfn_valid(pfn))
3454				continue;
3455			if (!early_pfn_in_nid(pfn, nid))
3456				continue;
3457		}
3458		page = pfn_to_page(pfn);
3459		set_page_links(page, zone, nid, pfn);
3460		mminit_verify_page_links(page, zone, nid, pfn);
3461		init_page_count(page);
3462		reset_page_mapcount(page);
3463		SetPageReserved(page);
3464		/*
3465		 * Mark the block movable so that blocks are reserved for
3466		 * movable at startup. This will force kernel allocations
3467		 * to reserve their blocks rather than leaking throughout
3468		 * the address space during boot when many long-lived
3469		 * kernel allocations are made. Later some blocks near
3470		 * the start are marked MIGRATE_RESERVE by
3471		 * setup_zone_migrate_reserve()
3472		 *
3473		 * bitmap is created for zone's valid pfn range. but memmap
3474		 * can be created for invalid pages (for alignment)
3475		 * check here not to call set_pageblock_migratetype() against
3476		 * pfn out of zone.
3477		 */
3478		if ((z->zone_start_pfn <= pfn)
3479		    && (pfn < z->zone_start_pfn + z->spanned_pages)
3480		    && !(pfn & (pageblock_nr_pages - 1)))
3481			set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3482
3483		INIT_LIST_HEAD(&page->lru);
3484#ifdef WANT_PAGE_VIRTUAL
3485		/* The shift won't overflow because ZONE_NORMAL is below 4G. */
3486		if (!is_highmem_idx(zone))
3487			set_page_address(page, __va(pfn << PAGE_SHIFT));
3488#endif
3489	}
3490}
3491
3492static void __meminit zone_init_free_lists(struct zone *zone)
3493{
3494	int order, t;
3495	for_each_migratetype_order(order, t) {
3496		INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3497		zone->free_area[order].nr_free = 0;
3498	}
3499}
3500
3501#ifndef __HAVE_ARCH_MEMMAP_INIT
3502#define memmap_init(size, nid, zone, start_pfn) \
3503	memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3504#endif
3505
3506static int zone_batchsize(struct zone *zone)
3507{
3508#ifdef CONFIG_MMU
3509	int batch;
3510
3511	/*
3512	 * The per-cpu-pages pools are set to around 1000th of the
3513	 * size of the zone.  But no more than 1/2 of a meg.
3514	 *
3515	 * OK, so we don't know how big the cache is.  So guess.
3516	 */
3517	batch = zone->present_pages / 1024;
3518	if (batch * PAGE_SIZE > 512 * 1024)
3519		batch = (512 * 1024) / PAGE_SIZE;
3520	batch /= 4;		/* We effectively *= 4 below */
3521	if (batch < 1)
3522		batch = 1;
3523
3524	/*
3525	 * Clamp the batch to a 2^n - 1 value. Having a power
3526	 * of 2 value was found to be more likely to have
3527	 * suboptimal cache aliasing properties in some cases.
3528	 *
3529	 * For example if 2 tasks are alternately allocating
3530	 * batches of pages, one task can end up with a lot
3531	 * of pages of one half of the possible page colors
3532	 * and the other with pages of the other colors.
3533	 */
3534	batch = rounddown_pow_of_two(batch + batch/2) - 1;
3535
3536	return batch;
3537
3538#else
3539	/* The deferral and batching of frees should be suppressed under NOMMU
3540	 * conditions.
3541	 *
3542	 * The problem is that NOMMU needs to be able to allocate large chunks
3543	 * of contiguous memory as there's no hardware page translation to
3544	 * assemble apparent contiguous memory from discontiguous pages.
3545	 *
3546	 * Queueing large contiguous runs of pages for batching, however,
3547	 * causes the pages to actually be freed in smaller chunks.  As there
3548	 * can be a significant delay between the individual batches being
3549	 * recycled, this leads to the once large chunks of space being
3550	 * fragmented and becoming unavailable for high-order allocations.
3551	 */
3552	return 0;
3553#endif
3554}
3555
3556static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3557{
3558	struct per_cpu_pages *pcp;
3559	int migratetype;
3560
3561	memset(p, 0, sizeof(*p));
3562
3563	pcp = &p->pcp;
3564	pcp->count = 0;
3565	pcp->high = 6 * batch;
3566	pcp->batch = max(1UL, 1 * batch);
3567	for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3568		INIT_LIST_HEAD(&pcp->lists[migratetype]);
3569}
3570
3571/*
3572 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3573 * to the value high for the pageset p.
3574 */
3575
3576static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3577				unsigned long high)
3578{
3579	struct per_cpu_pages *pcp;
3580
3581	pcp = &p->pcp;
3582	pcp->high = high;
3583	pcp->batch = max(1UL, high/4);
3584	if ((high/4) > (PAGE_SHIFT * 8))
3585		pcp->batch = PAGE_SHIFT * 8;
3586}
3587
3588static void setup_zone_pageset(struct zone *zone)
3589{
3590	int cpu;
3591
3592	zone->pageset = alloc_percpu(struct per_cpu_pageset);
3593
3594	for_each_possible_cpu(cpu) {
3595		struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3596
3597		setup_pageset(pcp, zone_batchsize(zone));
3598
3599		if (percpu_pagelist_fraction)
3600			setup_pagelist_highmark(pcp,
3601				(zone->present_pages /
3602					percpu_pagelist_fraction));
3603	}
3604}
3605
3606/*
3607 * Allocate per cpu pagesets and initialize them.
3608 * Before this call only boot pagesets were available.
3609 */
3610void __init setup_per_cpu_pageset(void)
3611{
3612	struct zone *zone;
3613
3614	for_each_populated_zone(zone)
3615		setup_zone_pageset(zone);
3616}
3617
3618static noinline __init_refok
3619int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3620{
3621	int i;
3622	struct pglist_data *pgdat = zone->zone_pgdat;
3623	size_t alloc_size;
3624
3625	/*
3626	 * The per-page waitqueue mechanism uses hashed waitqueues
3627	 * per zone.
3628	 */
3629	zone->wait_table_hash_nr_entries =
3630		 wait_table_hash_nr_entries(zone_size_pages);
3631	zone->wait_table_bits =
3632		wait_table_bits(zone->wait_table_hash_nr_entries);
3633	alloc_size = zone->wait_table_hash_nr_entries
3634					* sizeof(wait_queue_head_t);
3635
3636	if (!slab_is_available()) {
3637		zone->wait_table = (wait_queue_head_t *)
3638			alloc_bootmem_node_nopanic(pgdat, alloc_size);
3639	} else {
3640		/*
3641		 * This case means that a zone whose size was 0 gets new memory
3642		 * via memory hot-add.
3643		 * But it may be the case that a new node was hot-added.  In
3644		 * this case vmalloc() will not be able to use this new node's
3645		 * memory - this wait_table must be initialized to use this new
3646		 * node itself as well.
3647		 * To use this new node's memory, further consideration will be
3648		 * necessary.
3649		 */
3650		zone->wait_table = vmalloc(alloc_size);
3651	}
3652	if (!zone->wait_table)
3653		return -ENOMEM;
3654
3655	for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3656		init_waitqueue_head(zone->wait_table + i);
3657
3658	return 0;
3659}
3660
3661static int __zone_pcp_update(void *data)
3662{
3663	struct zone *zone = data;
3664	int cpu;
3665	unsigned long batch = zone_batchsize(zone), flags;
3666
3667	for_each_possible_cpu(cpu) {
3668		struct per_cpu_pageset *pset;
3669		struct per_cpu_pages *pcp;
3670
3671		pset = per_cpu_ptr(zone->pageset, cpu);
3672		pcp = &pset->pcp;
3673
3674		local_irq_save(flags);
3675		free_pcppages_bulk(zone, pcp->count, pcp);
3676		setup_pageset(pset, batch);
3677		local_irq_restore(flags);
3678	}
3679	return 0;
3680}
3681
3682void zone_pcp_update(struct zone *zone)
3683{
3684	stop_machine(__zone_pcp_update, zone, NULL);
3685}
3686
3687static __meminit void zone_pcp_init(struct zone *zone)
3688{
3689	/*
3690	 * per cpu subsystem is not up at this point. The following code
3691	 * relies on the ability of the linker to provide the
3692	 * offset of a (static) per cpu variable into the per cpu area.
 
 
 
3693	 */
3694	zone->pageset = &boot_pageset;
3695
3696	if (zone->present_pages)
3697		printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
3698			zone->name, zone->present_pages,
3699					 zone_batchsize(zone));
3700}
3701
3702__meminit int init_currently_empty_zone(struct zone *zone,
3703					unsigned long zone_start_pfn,
3704					unsigned long size,
3705					enum memmap_context context)
3706{
3707	struct pglist_data *pgdat = zone->zone_pgdat;
3708	int ret;
3709	ret = zone_wait_table_init(zone, size);
3710	if (ret)
3711		return ret;
3712	pgdat->nr_zones = zone_idx(zone) + 1;
3713
3714	zone->zone_start_pfn = zone_start_pfn;
3715
3716	mminit_dprintk(MMINIT_TRACE, "memmap_init",
3717			"Initialising map node %d zone %lu pfns %lu -> %lu\n",
3718			pgdat->node_id,
3719			(unsigned long)zone_idx(zone),
3720			zone_start_pfn, (zone_start_pfn + size));
3721
3722	zone_init_free_lists(zone);
3723
3724	return 0;
3725}
3726
3727#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3728/*
3729 * Basic iterator support. Return the first range of PFNs for a node
3730 * Note: nid == MAX_NUMNODES returns first region regardless of node
3731 */
3732static int __meminit first_active_region_index_in_nid(int nid)
3733{
3734	int i;
3735
3736	for (i = 0; i < nr_nodemap_entries; i++)
3737		if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3738			return i;
3739
3740	return -1;
3741}
3742
3743/*
3744 * Basic iterator support. Return the next active range of PFNs for a node
3745 * Note: nid == MAX_NUMNODES returns next region regardless of node
3746 */
3747static int __meminit next_active_region_index_in_nid(int index, int nid)
3748{
3749	for (index = index + 1; index < nr_nodemap_entries; index++)
3750		if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3751			return index;
3752
3753	return -1;
3754}
3755
3756#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3757/*
3758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3759 * Architectures may implement their own version but if add_active_range()
3760 * was used and there are no special requirements, this is a convenient
3761 * alternative
3762 */
3763int __meminit __early_pfn_to_nid(unsigned long pfn)
3764{
3765	int i;
3766
3767	for (i = 0; i < nr_nodemap_entries; i++) {
3768		unsigned long start_pfn = early_node_map[i].start_pfn;
3769		unsigned long end_pfn = early_node_map[i].end_pfn;
3770
3771		if (start_pfn <= pfn && pfn < end_pfn)
3772			return early_node_map[i].nid;
3773	}
3774	/* This is a memory hole */
3775	return -1;
3776}
3777#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3778
3779int __meminit early_pfn_to_nid(unsigned long pfn)
3780{
3781	int nid;
3782
3783	nid = __early_pfn_to_nid(pfn);
3784	if (nid >= 0)
3785		return nid;
3786	/* just returns 0 */
3787	return 0;
3788}
3789
3790#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3791bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3792{
3793	int nid;
3794
3795	nid = __early_pfn_to_nid(pfn);
3796	if (nid >= 0 && nid != node)
3797		return false;
3798	return true;
3799}
3800#endif
3801
3802/* Basic iterator support to walk early_node_map[] */
3803#define for_each_active_range_index_in_nid(i, nid) \
3804	for (i = first_active_region_index_in_nid(nid); i != -1; \
3805				i = next_active_region_index_in_nid(i, nid))
3806
3807/**
3808 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3809 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3810 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3811 *
3812 * If an architecture guarantees that all ranges registered with
3813 * add_active_ranges() contain no holes and may be freed, this
3814 * this function may be used instead of calling free_bootmem() manually.
3815 */
3816void __init free_bootmem_with_active_regions(int nid,
3817						unsigned long max_low_pfn)
3818{
3819	int i;
3820
3821	for_each_active_range_index_in_nid(i, nid) {
3822		unsigned long size_pages = 0;
3823		unsigned long end_pfn = early_node_map[i].end_pfn;
3824
3825		if (early_node_map[i].start_pfn >= max_low_pfn)
3826			continue;
3827
3828		if (end_pfn > max_low_pfn)
3829			end_pfn = max_low_pfn;
3830
3831		size_pages = end_pfn - early_node_map[i].start_pfn;
3832		free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3833				PFN_PHYS(early_node_map[i].start_pfn),
3834				size_pages << PAGE_SHIFT);
3835	}
3836}
3837
3838#ifdef CONFIG_HAVE_MEMBLOCK
3839/*
3840 * Basic iterator support. Return the last range of PFNs for a node
3841 * Note: nid == MAX_NUMNODES returns last region regardless of node
 
 
 
 
 
 
 
 
 
 
 
 
 
3842 */
3843static int __meminit last_active_region_index_in_nid(int nid)
 
3844{
3845	int i;
3846
3847	for (i = nr_nodemap_entries - 1; i >= 0; i--)
3848		if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3849			return i;
3850
3851	return -1;
3852}
3853
3854/*
3855 * Basic iterator support. Return the previous active range of PFNs for a node
3856 * Note: nid == MAX_NUMNODES returns next region regardless of node
3857 */
3858static int __meminit previous_active_region_index_in_nid(int index, int nid)
3859{
3860	for (index = index - 1; index >= 0; index--)
3861		if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3862			return index;
3863
3864	return -1;
3865}
3866
3867#define for_each_active_range_index_in_nid_reverse(i, nid) \
3868	for (i = last_active_region_index_in_nid(nid); i != -1; \
3869				i = previous_active_region_index_in_nid(i, nid))
3870
3871u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3872					u64 goal, u64 limit)
3873{
3874	int i;
3875
3876	/* Need to go over early_node_map to find out good range for node */
3877	for_each_active_range_index_in_nid_reverse(i, nid) {
3878		u64 addr;
3879		u64 ei_start, ei_last;
3880		u64 final_start, final_end;
3881
3882		ei_last = early_node_map[i].end_pfn;
3883		ei_last <<= PAGE_SHIFT;
3884		ei_start = early_node_map[i].start_pfn;
3885		ei_start <<= PAGE_SHIFT;
3886
3887		final_start = max(ei_start, goal);
3888		final_end = min(ei_last, limit);
3889
3890		if (final_start >= final_end)
3891			continue;
3892
3893		addr = memblock_find_in_range(final_start, final_end, size, align);
3894
3895		if (addr == MEMBLOCK_ERROR)
3896			continue;
3897
3898		return addr;
3899	}
3900
3901	return MEMBLOCK_ERROR;
3902}
3903#endif
3904
3905int __init add_from_early_node_map(struct range *range, int az,
3906				   int nr_range, int nid)
3907{
3908	int i;
3909	u64 start, end;
3910
3911	/* need to go over early_node_map to find out good range for node */
3912	for_each_active_range_index_in_nid(i, nid) {
3913		start = early_node_map[i].start_pfn;
3914		end = early_node_map[i].end_pfn;
3915		nr_range = add_range(range, az, nr_range, start, end);
3916	}
3917	return nr_range;
3918}
3919
3920void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3921{
3922	int i;
3923	int ret;
3924
3925	for_each_active_range_index_in_nid(i, nid) {
3926		ret = work_fn(early_node_map[i].start_pfn,
3927			      early_node_map[i].end_pfn, data);
3928		if (ret)
3929			break;
3930	}
3931}
3932/**
3933 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3934 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3935 *
3936 * If an architecture guarantees that all ranges registered with
3937 * add_active_ranges() contain no holes and may be freed, this
3938 * function may be used instead of calling memory_present() manually.
3939 */
3940void __init sparse_memory_present_with_active_regions(int nid)
3941{
3942	int i;
3943
3944	for_each_active_range_index_in_nid(i, nid)
3945		memory_present(early_node_map[i].nid,
3946				early_node_map[i].start_pfn,
3947				early_node_map[i].end_pfn);
3948}
3949
3950/**
3951 * get_pfn_range_for_nid - Return the start and end page frames for a node
3952 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3953 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3954 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3955 *
3956 * It returns the start and end page frame of a node based on information
3957 * provided by an arch calling add_active_range(). If called for a node
3958 * with no available memory, a warning is printed and the start and end
3959 * PFNs will be 0.
3960 */
3961void __meminit get_pfn_range_for_nid(unsigned int nid,
3962			unsigned long *start_pfn, unsigned long *end_pfn)
3963{
3964	int i;
3965	*start_pfn = -1UL;
3966	*end_pfn = 0;
3967
3968	for_each_active_range_index_in_nid(i, nid) {
3969		*start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3970		*end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3971	}
3972
3973	if (*start_pfn == -1UL)
3974		*start_pfn = 0;
3975}
3976
3977/*
3978 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3979 * assumption is made that zones within a node are ordered in monotonic
3980 * increasing memory addresses so that the "highest" populated zone is used
3981 */
3982static void __init find_usable_zone_for_movable(void)
3983{
3984	int zone_index;
3985	for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3986		if (zone_index == ZONE_MOVABLE)
3987			continue;
3988
3989		if (arch_zone_highest_possible_pfn[zone_index] >
3990				arch_zone_lowest_possible_pfn[zone_index])
3991			break;
3992	}
3993
3994	VM_BUG_ON(zone_index == -1);
3995	movable_zone = zone_index;
3996}
3997
3998/*
3999 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4000 * because it is sized independent of architecture. Unlike the other zones,
4001 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4002 * in each node depending on the size of each node and how evenly kernelcore
4003 * is distributed. This helper function adjusts the zone ranges
4004 * provided by the architecture for a given node by using the end of the
4005 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4006 * zones within a node are in order of monotonic increases memory addresses
4007 */
4008static void __meminit adjust_zone_range_for_zone_movable(int nid,
4009					unsigned long zone_type,
4010					unsigned long node_start_pfn,
4011					unsigned long node_end_pfn,
4012					unsigned long *zone_start_pfn,
4013					unsigned long *zone_end_pfn)
4014{
4015	/* Only adjust if ZONE_MOVABLE is on this node */
4016	if (zone_movable_pfn[nid]) {
4017		/* Size ZONE_MOVABLE */
4018		if (zone_type == ZONE_MOVABLE) {
4019			*zone_start_pfn = zone_movable_pfn[nid];
4020			*zone_end_pfn = min(node_end_pfn,
4021				arch_zone_highest_possible_pfn[movable_zone]);
4022
4023		/* Adjust for ZONE_MOVABLE starting within this range */
4024		} else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4025				*zone_end_pfn > zone_movable_pfn[nid]) {
4026			*zone_end_pfn = zone_movable_pfn[nid];
4027
4028		/* Check if this whole range is within ZONE_MOVABLE */
4029		} else if (*zone_start_pfn >= zone_movable_pfn[nid])
4030			*zone_start_pfn = *zone_end_pfn;
4031	}
4032}
4033
4034/*
4035 * Return the number of pages a zone spans in a node, including holes
4036 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4037 */
4038static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4039					unsigned long zone_type,
4040					unsigned long *ignored)
4041{
4042	unsigned long node_start_pfn, node_end_pfn;
4043	unsigned long zone_start_pfn, zone_end_pfn;
4044
4045	/* Get the start and end of the node and zone */
4046	get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4047	zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4048	zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4049	adjust_zone_range_for_zone_movable(nid, zone_type,
4050				node_start_pfn, node_end_pfn,
4051				&zone_start_pfn, &zone_end_pfn);
4052
4053	/* Check that this node has pages within the zone's required range */
4054	if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4055		return 0;
4056
4057	/* Move the zone boundaries inside the node if necessary */
4058	zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4059	zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4060
4061	/* Return the spanned pages */
4062	return zone_end_pfn - zone_start_pfn;
4063}
4064
4065/*
4066 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4067 * then all holes in the requested range will be accounted for.
4068 */
4069unsigned long __meminit __absent_pages_in_range(int nid,
4070				unsigned long range_start_pfn,
4071				unsigned long range_end_pfn)
4072{
4073	int i = 0;
4074	unsigned long prev_end_pfn = 0, hole_pages = 0;
4075	unsigned long start_pfn;
4076
4077	/* Find the end_pfn of the first active range of pfns in the node */
4078	i = first_active_region_index_in_nid(nid);
4079	if (i == -1)
4080		return 0;
4081
4082	prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4083
4084	/* Account for ranges before physical memory on this node */
4085	if (early_node_map[i].start_pfn > range_start_pfn)
4086		hole_pages = prev_end_pfn - range_start_pfn;
4087
4088	/* Find all holes for the zone within the node */
4089	for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4090
4091		/* No need to continue if prev_end_pfn is outside the zone */
4092		if (prev_end_pfn >= range_end_pfn)
4093			break;
4094
4095		/* Make sure the end of the zone is not within the hole */
4096		start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4097		prev_end_pfn = max(prev_end_pfn, range_start_pfn);
 
4098
4099		/* Update the hole size cound and move on */
4100		if (start_pfn > range_start_pfn) {
4101			BUG_ON(prev_end_pfn > start_pfn);
4102			hole_pages += start_pfn - prev_end_pfn;
4103		}
4104		prev_end_pfn = early_node_map[i].end_pfn;
4105	}
4106
4107	/* Account for ranges past physical memory on this node */
4108	if (range_end_pfn > prev_end_pfn)
4109		hole_pages += range_end_pfn -
4110				max(range_start_pfn, prev_end_pfn);
4111
4112	return hole_pages;
4113}
4114
4115/**
4116 * absent_pages_in_range - Return number of page frames in holes within a range
4117 * @start_pfn: The start PFN to start searching for holes
4118 * @end_pfn: The end PFN to stop searching for holes
4119 *
4120 * It returns the number of pages frames in memory holes within a range.
4121 */
4122unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4123							unsigned long end_pfn)
4124{
4125	return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4126}
4127
4128/* Return the number of page frames in holes in a zone on a node */
4129static unsigned long __meminit zone_absent_pages_in_node(int nid,
4130					unsigned long zone_type,
4131					unsigned long *ignored)
4132{
4133	unsigned long node_start_pfn, node_end_pfn;
4134	unsigned long zone_start_pfn, zone_end_pfn;
4135
4136	get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4137	zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4138							node_start_pfn);
4139	zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4140							node_end_pfn);
4141
4142	adjust_zone_range_for_zone_movable(nid, zone_type,
4143			node_start_pfn, node_end_pfn,
4144			&zone_start_pfn, &zone_end_pfn);
4145	return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4146}
4147
4148#else
4149static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4150					unsigned long zone_type,
4151					unsigned long *zones_size)
4152{
4153	return zones_size[zone_type];
4154}
4155
4156static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4157						unsigned long zone_type,
4158						unsigned long *zholes_size)
4159{
4160	if (!zholes_size)
4161		return 0;
4162
4163	return zholes_size[zone_type];
4164}
4165
4166#endif
4167
4168static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4169		unsigned long *zones_size, unsigned long *zholes_size)
4170{
4171	unsigned long realtotalpages, totalpages = 0;
4172	enum zone_type i;
4173
4174	for (i = 0; i < MAX_NR_ZONES; i++)
4175		totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4176								zones_size);
4177	pgdat->node_spanned_pages = totalpages;
4178
4179	realtotalpages = totalpages;
4180	for (i = 0; i < MAX_NR_ZONES; i++)
4181		realtotalpages -=
4182			zone_absent_pages_in_node(pgdat->node_id, i,
4183								zholes_size);
4184	pgdat->node_present_pages = realtotalpages;
4185	printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4186							realtotalpages);
4187}
4188
4189#ifndef CONFIG_SPARSEMEM
4190/*
4191 * Calculate the size of the zone->blockflags rounded to an unsigned long
4192 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4193 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4194 * round what is now in bits to nearest long in bits, then return it in
4195 * bytes.
4196 */
4197static unsigned long __init usemap_size(unsigned long zonesize)
4198{
4199	unsigned long usemapsize;
4200
4201	usemapsize = roundup(zonesize, pageblock_nr_pages);
4202	usemapsize = usemapsize >> pageblock_order;
4203	usemapsize *= NR_PAGEBLOCK_BITS;
4204	usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4205
4206	return usemapsize / 8;
4207}
4208
4209static void __init setup_usemap(struct pglist_data *pgdat,
4210				struct zone *zone, unsigned long zonesize)
4211{
4212	unsigned long usemapsize = usemap_size(zonesize);
4213	zone->pageblock_flags = NULL;
4214	if (usemapsize)
4215		zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4216								   usemapsize);
4217}
4218#else
4219static inline void setup_usemap(struct pglist_data *pgdat,
4220				struct zone *zone, unsigned long zonesize) {}
4221#endif /* CONFIG_SPARSEMEM */
4222
4223#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4224
4225/* Return a sensible default order for the pageblock size. */
4226static inline int pageblock_default_order(void)
4227{
4228	if (HPAGE_SHIFT > PAGE_SHIFT)
4229		return HUGETLB_PAGE_ORDER;
 
4230
4231	return MAX_ORDER-1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4232}
4233
4234/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4235static inline void __init set_pageblock_order(unsigned int order)
4236{
4237	/* Check that pageblock_nr_pages has not already been setup */
4238	if (pageblock_order)
4239		return;
4240
4241	/*
4242	 * Assume the largest contiguous order of interest is a huge page.
4243	 * This value may be variable depending on boot parameters on IA64
4244	 */
4245	pageblock_order = order;
4246}
4247#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4248
4249/*
4250 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4251 * and pageblock_default_order() are unused as pageblock_order is set
4252 * at compile-time. See include/linux/pageblock-flags.h for the values of
4253 * pageblock_order based on the kernel config
4254 */
4255static inline int pageblock_default_order(unsigned int order)
4256{
4257	return MAX_ORDER-1;
4258}
4259#define set_pageblock_order(x)	do {} while (0)
4260
4261#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4262
4263/*
4264 * Set up the zone data structures:
4265 *   - mark all pages reserved
4266 *   - mark all memory queues empty
4267 *   - clear the memory bitmaps
4268 */
4269static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4270		unsigned long *zones_size, unsigned long *zholes_size)
4271{
4272	enum zone_type j;
4273	int nid = pgdat->node_id;
4274	unsigned long zone_start_pfn = pgdat->node_start_pfn;
4275	int ret;
4276
4277	pgdat_resize_init(pgdat);
4278	pgdat->nr_zones = 0;
4279	init_waitqueue_head(&pgdat->kswapd_wait);
4280	pgdat->kswapd_max_order = 0;
4281	pgdat_page_cgroup_init(pgdat);
4282	
4283	for (j = 0; j < MAX_NR_ZONES; j++) {
4284		struct zone *zone = pgdat->node_zones + j;
4285		unsigned long size, realsize, memmap_pages;
4286		enum lru_list l;
4287
4288		size = zone_spanned_pages_in_node(nid, j, zones_size);
4289		realsize = size - zone_absent_pages_in_node(nid, j,
4290								zholes_size);
4291
4292		/*
4293		 * Adjust realsize so that it accounts for how much memory
4294		 * is used by this zone for memmap. This affects the watermark
4295		 * and per-cpu initialisations
4296		 */
4297		memmap_pages =
4298			PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4299		if (realsize >= memmap_pages) {
4300			realsize -= memmap_pages;
4301			if (memmap_pages)
4302				printk(KERN_DEBUG
4303				       "  %s zone: %lu pages used for memmap\n",
4304				       zone_names[j], memmap_pages);
4305		} else
4306			printk(KERN_WARNING
4307				"  %s zone: %lu pages exceeds realsize %lu\n",
4308				zone_names[j], memmap_pages, realsize);
4309
4310		/* Account for reserved pages */
4311		if (j == 0 && realsize > dma_reserve) {
4312			realsize -= dma_reserve;
4313			printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
4314					zone_names[0], dma_reserve);
4315		}
4316
4317		if (!is_highmem_idx(j))
4318			nr_kernel_pages += realsize;
4319		nr_all_pages += realsize;
4320
4321		zone->spanned_pages = size;
4322		zone->present_pages = realsize;
4323#ifdef CONFIG_NUMA
4324		zone->node = nid;
4325		zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4326						/ 100;
4327		zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4328#endif
4329		zone->name = zone_names[j];
4330		spin_lock_init(&zone->lock);
4331		spin_lock_init(&zone->lru_lock);
4332		zone_seqlock_init(zone);
4333		zone->zone_pgdat = pgdat;
4334
4335		zone_pcp_init(zone);
4336		for_each_lru(l)
4337			INIT_LIST_HEAD(&zone->lru[l].list);
4338		zone->reclaim_stat.recent_rotated[0] = 0;
4339		zone->reclaim_stat.recent_rotated[1] = 0;
4340		zone->reclaim_stat.recent_scanned[0] = 0;
4341		zone->reclaim_stat.recent_scanned[1] = 0;
4342		zap_zone_vm_stats(zone);
4343		zone->flags = 0;
4344		if (!size)
4345			continue;
4346
4347		set_pageblock_order(pageblock_default_order());
4348		setup_usemap(pgdat, zone, size);
4349		ret = init_currently_empty_zone(zone, zone_start_pfn,
4350						size, MEMMAP_EARLY);
4351		BUG_ON(ret);
4352		memmap_init(size, nid, j, zone_start_pfn);
4353		zone_start_pfn += size;
4354	}
4355}
4356
4357static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4358{
4359	/* Skip empty nodes */
4360	if (!pgdat->node_spanned_pages)
4361		return;
4362
4363#ifdef CONFIG_FLAT_NODE_MEM_MAP
4364	/* ia64 gets its own node_mem_map, before this, without bootmem */
4365	if (!pgdat->node_mem_map) {
4366		unsigned long size, start, end;
4367		struct page *map;
4368
4369		/*
4370		 * The zone's endpoints aren't required to be MAX_ORDER
4371		 * aligned but the node_mem_map endpoints must be in order
4372		 * for the buddy allocator to function correctly.
4373		 */
4374		start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4375		end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4376		end = ALIGN(end, MAX_ORDER_NR_PAGES);
4377		size =  (end - start) * sizeof(struct page);
4378		map = alloc_remap(pgdat->node_id, size);
4379		if (!map)
4380			map = alloc_bootmem_node_nopanic(pgdat, size);
4381		pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4382	}
4383#ifndef CONFIG_NEED_MULTIPLE_NODES
4384	/*
4385	 * With no DISCONTIG, the global mem_map is just set as node 0's
 
 
 
4386	 */
4387	if (pgdat == NODE_DATA(0)) {
4388		mem_map = NODE_DATA(0)->node_mem_map;
4389#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4390		if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4391			mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4392#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4393	}
4394#endif
4395#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4396}
4397
4398void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4399		unsigned long node_start_pfn, unsigned long *zholes_size)
4400{
4401	pg_data_t *pgdat = NODE_DATA(nid);
4402
4403	pgdat->node_id = nid;
4404	pgdat->node_start_pfn = node_start_pfn;
4405	calculate_node_totalpages(pgdat, zones_size, zholes_size);
4406
4407	alloc_node_mem_map(pgdat);
4408#ifdef CONFIG_FLAT_NODE_MEM_MAP
4409	printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4410		nid, (unsigned long)pgdat,
4411		(unsigned long)pgdat->node_mem_map);
4412#endif
4413
4414	free_area_init_core(pgdat, zones_size, zholes_size);
4415}
4416
4417#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4418
4419#if MAX_NUMNODES > 1
4420/*
4421 * Figure out the number of possible node ids.
4422 */
4423static void __init setup_nr_node_ids(void)
4424{
4425	unsigned int node;
4426	unsigned int highest = 0;
4427
4428	for_each_node_mask(node, node_possible_map)
4429		highest = node;
4430	nr_node_ids = highest + 1;
4431}
4432#else
4433static inline void setup_nr_node_ids(void)
4434{
4435}
4436#endif
4437
4438/**
4439 * add_active_range - Register a range of PFNs backed by physical memory
4440 * @nid: The node ID the range resides on
4441 * @start_pfn: The start PFN of the available physical memory
4442 * @end_pfn: The end PFN of the available physical memory
4443 *
4444 * These ranges are stored in an early_node_map[] and later used by
4445 * free_area_init_nodes() to calculate zone sizes and holes. If the
4446 * range spans a memory hole, it is up to the architecture to ensure
4447 * the memory is not freed by the bootmem allocator. If possible
4448 * the range being registered will be merged with existing ranges.
4449 */
4450void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4451						unsigned long end_pfn)
4452{
4453	int i;
4454
4455	mminit_dprintk(MMINIT_TRACE, "memory_register",
4456			"Entering add_active_range(%d, %#lx, %#lx) "
4457			"%d entries of %d used\n",
4458			nid, start_pfn, end_pfn,
4459			nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4460
4461	mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4462
4463	/* Merge with existing active regions if possible */
4464	for (i = 0; i < nr_nodemap_entries; i++) {
4465		if (early_node_map[i].nid != nid)
4466			continue;
4467
4468		/* Skip if an existing region covers this new one */
4469		if (start_pfn >= early_node_map[i].start_pfn &&
4470				end_pfn <= early_node_map[i].end_pfn)
4471			return;
4472
4473		/* Merge forward if suitable */
4474		if (start_pfn <= early_node_map[i].end_pfn &&
4475				end_pfn > early_node_map[i].end_pfn) {
4476			early_node_map[i].end_pfn = end_pfn;
4477			return;
4478		}
4479
4480		/* Merge backward if suitable */
4481		if (start_pfn < early_node_map[i].start_pfn &&
4482				end_pfn >= early_node_map[i].start_pfn) {
4483			early_node_map[i].start_pfn = start_pfn;
4484			return;
4485		}
4486	}
4487
4488	/* Check that early_node_map is large enough */
4489	if (i >= MAX_ACTIVE_REGIONS) {
4490		printk(KERN_CRIT "More than %d memory regions, truncating\n",
4491							MAX_ACTIVE_REGIONS);
4492		return;
4493	}
4494
4495	early_node_map[i].nid = nid;
4496	early_node_map[i].start_pfn = start_pfn;
4497	early_node_map[i].end_pfn = end_pfn;
4498	nr_nodemap_entries = i + 1;
4499}
4500
4501/**
4502 * remove_active_range - Shrink an existing registered range of PFNs
4503 * @nid: The node id the range is on that should be shrunk
4504 * @start_pfn: The new PFN of the range
4505 * @end_pfn: The new PFN of the range
4506 *
4507 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4508 * The map is kept near the end physical page range that has already been
4509 * registered. This function allows an arch to shrink an existing registered
4510 * range.
4511 */
4512void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4513				unsigned long end_pfn)
4514{
4515	int i, j;
4516	int removed = 0;
4517
4518	printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4519			  nid, start_pfn, end_pfn);
4520
4521	/* Find the old active region end and shrink */
4522	for_each_active_range_index_in_nid(i, nid) {
4523		if (early_node_map[i].start_pfn >= start_pfn &&
4524		    early_node_map[i].end_pfn <= end_pfn) {
4525			/* clear it */
4526			early_node_map[i].start_pfn = 0;
4527			early_node_map[i].end_pfn = 0;
4528			removed = 1;
4529			continue;
4530		}
4531		if (early_node_map[i].start_pfn < start_pfn &&
4532		    early_node_map[i].end_pfn > start_pfn) {
4533			unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4534			early_node_map[i].end_pfn = start_pfn;
4535			if (temp_end_pfn > end_pfn)
4536				add_active_range(nid, end_pfn, temp_end_pfn);
4537			continue;
4538		}
4539		if (early_node_map[i].start_pfn >= start_pfn &&
4540		    early_node_map[i].end_pfn > end_pfn &&
4541		    early_node_map[i].start_pfn < end_pfn) {
4542			early_node_map[i].start_pfn = end_pfn;
4543			continue;
4544		}
4545	}
4546
4547	if (!removed)
4548		return;
4549
4550	/* remove the blank ones */
4551	for (i = nr_nodemap_entries - 1; i > 0; i--) {
4552		if (early_node_map[i].nid != nid)
4553			continue;
4554		if (early_node_map[i].end_pfn)
4555			continue;
4556		/* we found it, get rid of it */
4557		for (j = i; j < nr_nodemap_entries - 1; j++)
4558			memcpy(&early_node_map[j], &early_node_map[j+1],
4559				sizeof(early_node_map[j]));
4560		j = nr_nodemap_entries - 1;
4561		memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4562		nr_nodemap_entries--;
4563	}
4564}
4565
4566/**
4567 * remove_all_active_ranges - Remove all currently registered regions
4568 *
4569 * During discovery, it may be found that a table like SRAT is invalid
4570 * and an alternative discovery method must be used. This function removes
4571 * all currently registered regions.
4572 */
4573void __init remove_all_active_ranges(void)
4574{
4575	memset(early_node_map, 0, sizeof(early_node_map));
4576	nr_nodemap_entries = 0;
4577}
4578
4579/* Compare two active node_active_regions */
4580static int __init cmp_node_active_region(const void *a, const void *b)
4581{
4582	struct node_active_region *arange = (struct node_active_region *)a;
4583	struct node_active_region *brange = (struct node_active_region *)b;
4584
4585	/* Done this way to avoid overflows */
4586	if (arange->start_pfn > brange->start_pfn)
4587		return 1;
4588	if (arange->start_pfn < brange->start_pfn)
4589		return -1;
 
 
4590
4591	return 0;
 
 
4592}
4593
4594/* sort the node_map by start_pfn */
4595void __init sort_node_map(void)
4596{
4597	sort(early_node_map, (size_t)nr_nodemap_entries,
4598			sizeof(struct node_active_region),
4599			cmp_node_active_region, NULL);
 
 
 
4600}
 
4601
4602/**
4603 * node_map_pfn_alignment - determine the maximum internode alignment
4604 *
4605 * This function should be called after node map is populated and sorted.
4606 * It calculates the maximum power of two alignment which can distinguish
4607 * all the nodes.
4608 *
4609 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4610 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
4611 * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
4612 * shifted, 1GiB is enough and this function will indicate so.
4613 *
4614 * This is used to test whether pfn -> nid mapping of the chosen memory
4615 * model has fine enough granularity to avoid incorrect mapping for the
4616 * populated node map.
4617 *
4618 * Returns the determined alignment in pfn's.  0 if there is no alignment
4619 * requirement (single node).
4620 */
4621unsigned long __init node_map_pfn_alignment(void)
4622{
4623	unsigned long accl_mask = 0, last_end = 0;
4624	int last_nid = -1;
4625	int i;
4626
4627	for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4628		int nid = early_node_map[i].nid;
4629		unsigned long start = early_node_map[i].start_pfn;
4630		unsigned long end = early_node_map[i].end_pfn;
4631		unsigned long mask;
4632
4633		if (!start || last_nid < 0 || last_nid == nid) {
4634			last_nid = nid;
4635			last_end = end;
4636			continue;
4637		}
4638
4639		/*
4640		 * Start with a mask granular enough to pin-point to the
4641		 * start pfn and tick off bits one-by-one until it becomes
4642		 * too coarse to separate the current node from the last.
 
 
4643		 */
4644		mask = ~((1 << __ffs(start)) - 1);
4645		while (mask && last_end <= (start & (mask << 1)))
4646			mask <<= 1;
4647
4648		/* accumulate all internode masks */
4649		accl_mask |= mask;
4650	}
4651
4652	/* convert mask to number of pages */
4653	return ~accl_mask + 1;
4654}
4655
4656/* Find the lowest pfn for a node */
4657static unsigned long __init find_min_pfn_for_node(int nid)
4658{
4659	int i;
4660	unsigned long min_pfn = ULONG_MAX;
4661
4662	/* Assuming a sorted map, the first range found has the starting pfn */
4663	for_each_active_range_index_in_nid(i, nid)
4664		min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4665
4666	if (min_pfn == ULONG_MAX) {
4667		printk(KERN_WARNING
4668			"Could not find start_pfn for node %d\n", nid);
4669		return 0;
4670	}
4671
4672	return min_pfn;
4673}
4674
4675/**
4676 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4677 *
4678 * It returns the minimum PFN based on information provided via
4679 * add_active_range().
4680 */
4681unsigned long __init find_min_pfn_with_active_regions(void)
4682{
4683	return find_min_pfn_for_node(MAX_NUMNODES);
4684}
4685
4686/*
4687 * early_calculate_totalpages()
4688 * Sum pages in active regions for movable zone.
4689 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4690 */
4691static unsigned long __init early_calculate_totalpages(void)
4692{
4693	int i;
4694	unsigned long totalpages = 0;
4695
4696	for (i = 0; i < nr_nodemap_entries; i++) {
4697		unsigned long pages = early_node_map[i].end_pfn -
4698						early_node_map[i].start_pfn;
4699		totalpages += pages;
4700		if (pages)
4701			node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4702	}
4703  	return totalpages;
4704}
4705
4706/*
4707 * Find the PFN the Movable zone begins in each node. Kernel memory
4708 * is spread evenly between nodes as long as the nodes have enough
4709 * memory. When they don't, some nodes will have more kernelcore than
4710 * others
4711 */
4712static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4713{
4714	int i, nid;
4715	unsigned long usable_startpfn;
4716	unsigned long kernelcore_node, kernelcore_remaining;
4717	/* save the state before borrow the nodemask */
4718	nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4719	unsigned long totalpages = early_calculate_totalpages();
4720	int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4721
4722	/*
4723	 * If movablecore was specified, calculate what size of
4724	 * kernelcore that corresponds so that memory usable for
4725	 * any allocation type is evenly spread. If both kernelcore
4726	 * and movablecore are specified, then the value of kernelcore
4727	 * will be used for required_kernelcore if it's greater than
4728	 * what movablecore would have allowed.
4729	 */
4730	if (required_movablecore) {
4731		unsigned long corepages;
4732
4733		/*
4734		 * Round-up so that ZONE_MOVABLE is at least as large as what
4735		 * was requested by the user
4736		 */
4737		required_movablecore =
4738			roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4739		corepages = totalpages - required_movablecore;
4740
4741		required_kernelcore = max(required_kernelcore, corepages);
4742	}
4743
4744	/* If kernelcore was not specified, there is no ZONE_MOVABLE */
4745	if (!required_kernelcore)
4746		goto out;
4747
4748	/* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4749	find_usable_zone_for_movable();
4750	usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4751
4752restart:
4753	/* Spread kernelcore memory as evenly as possible throughout nodes */
4754	kernelcore_node = required_kernelcore / usable_nodes;
4755	for_each_node_state(nid, N_HIGH_MEMORY) {
4756		/*
4757		 * Recalculate kernelcore_node if the division per node
4758		 * now exceeds what is necessary to satisfy the requested
4759		 * amount of memory for the kernel
4760		 */
4761		if (required_kernelcore < kernelcore_node)
4762			kernelcore_node = required_kernelcore / usable_nodes;
4763
4764		/*
4765		 * As the map is walked, we track how much memory is usable
4766		 * by the kernel using kernelcore_remaining. When it is
4767		 * 0, the rest of the node is usable by ZONE_MOVABLE
4768		 */
4769		kernelcore_remaining = kernelcore_node;
4770
4771		/* Go through each range of PFNs within this node */
4772		for_each_active_range_index_in_nid(i, nid) {
4773			unsigned long start_pfn, end_pfn;
4774			unsigned long size_pages;
4775
4776			start_pfn = max(early_node_map[i].start_pfn,
4777						zone_movable_pfn[nid]);
4778			end_pfn = early_node_map[i].end_pfn;
4779			if (start_pfn >= end_pfn)
4780				continue;
4781
4782			/* Account for what is only usable for kernelcore */
4783			if (start_pfn < usable_startpfn) {
4784				unsigned long kernel_pages;
4785				kernel_pages = min(end_pfn, usable_startpfn)
4786								- start_pfn;
4787
4788				kernelcore_remaining -= min(kernel_pages,
4789							kernelcore_remaining);
4790				required_kernelcore -= min(kernel_pages,
4791							required_kernelcore);
4792
4793				/* Continue if range is now fully accounted */
4794				if (end_pfn <= usable_startpfn) {
4795
4796					/*
4797					 * Push zone_movable_pfn to the end so
4798					 * that if we have to rebalance
4799					 * kernelcore across nodes, we will
4800					 * not double account here
4801					 */
4802					zone_movable_pfn[nid] = end_pfn;
4803					continue;
4804				}
4805				start_pfn = usable_startpfn;
4806			}
4807
4808			/*
4809			 * The usable PFN range for ZONE_MOVABLE is from
4810			 * start_pfn->end_pfn. Calculate size_pages as the
4811			 * number of pages used as kernelcore
4812			 */
4813			size_pages = end_pfn - start_pfn;
4814			if (size_pages > kernelcore_remaining)
4815				size_pages = kernelcore_remaining;
4816			zone_movable_pfn[nid] = start_pfn + size_pages;
4817
4818			/*
4819			 * Some kernelcore has been met, update counts and
4820			 * break if the kernelcore for this node has been
4821			 * satisified
4822			 */
4823			required_kernelcore -= min(required_kernelcore,
4824								size_pages);
4825			kernelcore_remaining -= size_pages;
4826			if (!kernelcore_remaining)
4827				break;
4828		}
4829	}
4830
4831	/*
4832	 * If there is still required_kernelcore, we do another pass with one
4833	 * less node in the count. This will push zone_movable_pfn[nid] further
4834	 * along on the nodes that still have memory until kernelcore is
4835	 * satisified
 
4836	 */
4837	usable_nodes--;
4838	if (usable_nodes && required_kernelcore > usable_nodes)
4839		goto restart;
4840
4841	/* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4842	for (nid = 0; nid < MAX_NUMNODES; nid++)
4843		zone_movable_pfn[nid] =
4844			roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4845
4846out:
4847	/* restore the node_state */
4848	node_states[N_HIGH_MEMORY] = saved_node_state;
4849}
4850
4851/* Any regular memory on that node ? */
4852static void check_for_regular_memory(pg_data_t *pgdat)
4853{
4854#ifdef CONFIG_HIGHMEM
4855	enum zone_type zone_type;
4856
4857	for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4858		struct zone *zone = &pgdat->node_zones[zone_type];
4859		if (zone->present_pages)
4860			node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4861	}
4862#endif
4863}
4864
4865/**
4866 * free_area_init_nodes - Initialise all pg_data_t and zone data
4867 * @max_zone_pfn: an array of max PFNs for each zone
4868 *
4869 * This will call free_area_init_node() for each active node in the system.
4870 * Using the page ranges provided by add_active_range(), the size of each
4871 * zone in each node and their holes is calculated. If the maximum PFN
4872 * between two adjacent zones match, it is assumed that the zone is empty.
4873 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4874 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4875 * starts where the previous one ended. For example, ZONE_DMA32 starts
4876 * at arch_max_dma_pfn.
4877 */
4878void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4879{
4880	unsigned long nid;
4881	int i;
4882
4883	/* Sort early_node_map as initialisation assumes it is sorted */
4884	sort_node_map();
4885
4886	/* Record where the zone boundaries are */
4887	memset(arch_zone_lowest_possible_pfn, 0,
4888				sizeof(arch_zone_lowest_possible_pfn));
4889	memset(arch_zone_highest_possible_pfn, 0,
4890				sizeof(arch_zone_highest_possible_pfn));
4891	arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4892	arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4893	for (i = 1; i < MAX_NR_ZONES; i++) {
4894		if (i == ZONE_MOVABLE)
4895			continue;
4896		arch_zone_lowest_possible_pfn[i] =
4897			arch_zone_highest_possible_pfn[i-1];
4898		arch_zone_highest_possible_pfn[i] =
4899			max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900	}
4901	arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4902	arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903
4904	/* Find the PFNs that ZONE_MOVABLE begins at in each node */
4905	memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4906	find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4907
4908	/* Print out the zone ranges */
4909	printk("Zone PFN ranges:\n");
4910	for (i = 0; i < MAX_NR_ZONES; i++) {
4911		if (i == ZONE_MOVABLE)
4912			continue;
4913		printk("  %-8s ", zone_names[i]);
4914		if (arch_zone_lowest_possible_pfn[i] ==
4915				arch_zone_highest_possible_pfn[i])
4916			printk("empty\n");
4917		else
4918			printk("%0#10lx -> %0#10lx\n",
4919				arch_zone_lowest_possible_pfn[i],
4920				arch_zone_highest_possible_pfn[i]);
4921	}
4922
4923	/* Print out the PFNs ZONE_MOVABLE begins at in each node */
4924	printk("Movable zone start PFN for each node\n");
4925	for (i = 0; i < MAX_NUMNODES; i++) {
4926		if (zone_movable_pfn[i])
4927			printk("  Node %d: %lu\n", i, zone_movable_pfn[i]);
4928	}
4929
4930	/* Print out the early_node_map[] */
4931	printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4932	for (i = 0; i < nr_nodemap_entries; i++)
4933		printk("  %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4934						early_node_map[i].start_pfn,
4935						early_node_map[i].end_pfn);
4936
4937	/* Initialise every node */
4938	mminit_verify_pageflags_layout();
4939	setup_nr_node_ids();
4940	for_each_online_node(nid) {
4941		pg_data_t *pgdat = NODE_DATA(nid);
4942		free_area_init_node(nid, NULL,
4943				find_min_pfn_for_node(nid), NULL);
4944
4945		/* Any memory on that node */
4946		if (pgdat->node_present_pages)
4947			node_set_state(nid, N_HIGH_MEMORY);
4948		check_for_regular_memory(pgdat);
4949	}
4950}
4951
4952static int __init cmdline_parse_core(char *p, unsigned long *core)
4953{
4954	unsigned long long coremem;
4955	if (!p)
4956		return -EINVAL;
4957
4958	coremem = memparse(p, &p);
4959	*core = coremem >> PAGE_SHIFT;
4960
4961	/* Paranoid check that UL is enough for the coremem value */
4962	WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4963
4964	return 0;
4965}
4966
4967/*
4968 * kernelcore=size sets the amount of memory for use for allocations that
4969 * cannot be reclaimed or migrated.
4970 */
4971static int __init cmdline_parse_kernelcore(char *p)
4972{
4973	return cmdline_parse_core(p, &required_kernelcore);
4974}
4975
4976/*
4977 * movablecore=size sets the amount of memory for use for allocations that
4978 * can be reclaimed or migrated.
4979 */
4980static int __init cmdline_parse_movablecore(char *p)
4981{
4982	return cmdline_parse_core(p, &required_movablecore);
4983}
4984
4985early_param("kernelcore", cmdline_parse_kernelcore);
4986early_param("movablecore", cmdline_parse_movablecore);
4987
4988#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4989
4990/**
4991 * set_dma_reserve - set the specified number of pages reserved in the first zone
4992 * @new_dma_reserve: The number of pages to mark reserved
4993 *
4994 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4995 * In the DMA zone, a significant percentage may be consumed by kernel image
4996 * and other unfreeable allocations which can skew the watermarks badly. This
4997 * function may optionally be used to account for unfreeable pages in the
4998 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4999 * smaller per-cpu batchsize.
5000 */
5001void __init set_dma_reserve(unsigned long new_dma_reserve)
5002{
5003	dma_reserve = new_dma_reserve;
5004}
5005
5006void __init free_area_init(unsigned long *zones_size)
5007{
5008	free_area_init_node(0, zones_size,
5009			__pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5010}
5011
5012static int page_alloc_cpu_notify(struct notifier_block *self,
5013				 unsigned long action, void *hcpu)
5014{
5015	int cpu = (unsigned long)hcpu;
5016
5017	if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5018		drain_pages(cpu);
5019
5020		/*
5021		 * Spill the event counters of the dead processor
5022		 * into the current processors event counters.
5023		 * This artificially elevates the count of the current
5024		 * processor.
5025		 */
5026		vm_events_fold_cpu(cpu);
5027
5028		/*
5029		 * Zero the differential counters of the dead processor
5030		 * so that the vm statistics are consistent.
5031		 *
5032		 * This is only okay since the processor is dead and cannot
5033		 * race with what we are doing.
5034		 */
5035		refresh_cpu_vm_stats(cpu);
5036	}
5037	return NOTIFY_OK;
5038}
5039
5040void __init page_alloc_init(void)
5041{
5042	hotcpu_notifier(page_alloc_cpu_notify, 0);
 
 
5043}
5044
5045/*
5046 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5047 *	or min_free_kbytes changes.
5048 */
5049static void calculate_totalreserve_pages(void)
5050{
5051	struct pglist_data *pgdat;
5052	unsigned long reserve_pages = 0;
5053	enum zone_type i, j;
5054
5055	for_each_online_pgdat(pgdat) {
 
 
 
5056		for (i = 0; i < MAX_NR_ZONES; i++) {
5057			struct zone *zone = pgdat->node_zones + i;
5058			unsigned long max = 0;
 
5059
5060			/* Find valid and maximum lowmem_reserve in the zone */
5061			for (j = i; j < MAX_NR_ZONES; j++) {
5062				if (zone->lowmem_reserve[j] > max)
5063					max = zone->lowmem_reserve[j];
5064			}
5065
5066			/* we treat the high watermark as reserved pages. */
5067			max += high_wmark_pages(zone);
5068
5069			if (max > zone->present_pages)
5070				max = zone->present_pages;
 
 
 
5071			reserve_pages += max;
5072		}
5073	}
5074	totalreserve_pages = reserve_pages;
5075}
5076
5077/*
5078 * setup_per_zone_lowmem_reserve - called whenever
5079 *	sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
5080 *	has a correct pages reserved value, so an adequate number of
5081 *	pages are left in the zone after a successful __alloc_pages().
5082 */
5083static void setup_per_zone_lowmem_reserve(void)
5084{
5085	struct pglist_data *pgdat;
5086	enum zone_type j, idx;
5087
5088	for_each_online_pgdat(pgdat) {
5089		for (j = 0; j < MAX_NR_ZONES; j++) {
5090			struct zone *zone = pgdat->node_zones + j;
5091			unsigned long present_pages = zone->present_pages;
5092
5093			zone->lowmem_reserve[j] = 0;
5094
5095			idx = j;
5096			while (idx) {
5097				struct zone *lower_zone;
5098
5099				idx--;
5100
5101				if (sysctl_lowmem_reserve_ratio[idx] < 1)
5102					sysctl_lowmem_reserve_ratio[idx] = 1;
5103
5104				lower_zone = pgdat->node_zones + idx;
5105				lower_zone->lowmem_reserve[j] = present_pages /
5106					sysctl_lowmem_reserve_ratio[idx];
5107				present_pages += lower_zone->present_pages;
5108			}
5109		}
5110	}
5111
5112	/* update totalreserve_pages */
5113	calculate_totalreserve_pages();
5114}
5115
5116/**
5117 * setup_per_zone_wmarks - called when min_free_kbytes changes
5118 * or when memory is hot-{added|removed}
5119 *
5120 * Ensures that the watermark[min,low,high] values for each zone are set
5121 * correctly with respect to min_free_kbytes.
5122 */
5123void setup_per_zone_wmarks(void)
5124{
5125	unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5126	unsigned long lowmem_pages = 0;
5127	struct zone *zone;
5128	unsigned long flags;
5129
5130	/* Calculate total number of !ZONE_HIGHMEM pages */
5131	for_each_zone(zone) {
5132		if (!is_highmem(zone))
5133			lowmem_pages += zone->present_pages;
5134	}
5135
5136	for_each_zone(zone) {
5137		u64 tmp;
5138
5139		spin_lock_irqsave(&zone->lock, flags);
5140		tmp = (u64)pages_min * zone->present_pages;
5141		do_div(tmp, lowmem_pages);
5142		if (is_highmem(zone)) {
5143			/*
5144			 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5145			 * need highmem pages, so cap pages_min to a small
5146			 * value here.
5147			 *
5148			 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5149			 * deltas controls asynch page reclaim, and so should
5150			 * not be capped for highmem.
5151			 */
5152			int min_pages;
5153
5154			min_pages = zone->present_pages / 1024;
5155			if (min_pages < SWAP_CLUSTER_MAX)
5156				min_pages = SWAP_CLUSTER_MAX;
5157			if (min_pages > 128)
5158				min_pages = 128;
5159			zone->watermark[WMARK_MIN] = min_pages;
5160		} else {
5161			/*
5162			 * If it's a lowmem zone, reserve a number of pages
5163			 * proportionate to the zone's size.
5164			 */
5165			zone->watermark[WMARK_MIN] = tmp;
5166		}
5167
5168		zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + (tmp >> 2);
5169		zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5170		setup_zone_migrate_reserve(zone);
 
 
 
 
 
 
 
 
 
 
 
5171		spin_unlock_irqrestore(&zone->lock, flags);
5172	}
5173
5174	/* update totalreserve_pages */
5175	calculate_totalreserve_pages();
5176}
5177
5178/*
5179 * The inactive anon list should be small enough that the VM never has to
5180 * do too much work, but large enough that each inactive page has a chance
5181 * to be referenced again before it is swapped out.
5182 *
5183 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5184 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5185 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5186 * the anonymous pages are kept on the inactive list.
5187 *
5188 * total     target    max
5189 * memory    ratio     inactive anon
5190 * -------------------------------------
5191 *   10MB       1         5MB
5192 *  100MB       1        50MB
5193 *    1GB       3       250MB
5194 *   10GB      10       0.9GB
5195 *  100GB      31         3GB
5196 *    1TB     101        10GB
5197 *   10TB     320        32GB
5198 */
5199static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5200{
5201	unsigned int gb, ratio;
5202
5203	/* Zone size in gigabytes */
5204	gb = zone->present_pages >> (30 - PAGE_SHIFT);
5205	if (gb)
5206		ratio = int_sqrt(10 * gb);
5207	else
5208		ratio = 1;
5209
5210	zone->inactive_ratio = ratio;
5211}
5212
5213static void __meminit setup_per_zone_inactive_ratio(void)
5214{
5215	struct zone *zone;
 
 
 
 
 
5216
 
 
 
 
5217	for_each_zone(zone)
5218		calculate_zone_inactive_ratio(zone);
5219}
5220
5221/*
5222 * Initialise min_free_kbytes.
5223 *
5224 * For small machines we want it small (128k min).  For large machines
5225 * we want it large (64MB max).  But it is not linear, because network
5226 * bandwidth does not increase linearly with machine size.  We use
5227 *
5228 * 	min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5229 *	min_free_kbytes = sqrt(lowmem_kbytes * 16)
5230 *
5231 * which yields
5232 *
5233 * 16MB:	512k
5234 * 32MB:	724k
5235 * 64MB:	1024k
5236 * 128MB:	1448k
5237 * 256MB:	2048k
5238 * 512MB:	2896k
5239 * 1024MB:	4096k
5240 * 2048MB:	5792k
5241 * 4096MB:	8192k
5242 * 8192MB:	11584k
5243 * 16384MB:	16384k
5244 */
5245int __meminit init_per_zone_wmark_min(void)
5246{
5247	unsigned long lowmem_kbytes;
 
5248
5249	lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
 
 
 
 
 
 
 
 
 
5250
5251	min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5252	if (min_free_kbytes < 128)
5253		min_free_kbytes = 128;
5254	if (min_free_kbytes > 65536)
5255		min_free_kbytes = 65536;
5256	setup_per_zone_wmarks();
5257	refresh_zone_stat_thresholds();
5258	setup_per_zone_lowmem_reserve();
5259	setup_per_zone_inactive_ratio();
 
 
 
 
 
 
 
5260	return 0;
5261}
5262module_init(init_per_zone_wmark_min)
5263
5264/*
5265 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
5266 *	that we can call two helper functions whenever min_free_kbytes
5267 *	changes.
5268 */
5269int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
5270	void __user *buffer, size_t *length, loff_t *ppos)
5271{
5272	proc_dointvec(table, write, buffer, length, ppos);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5273	if (write)
5274		setup_per_zone_wmarks();
 
5275	return 0;
5276}
5277
5278#ifdef CONFIG_NUMA
5279int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5280	void __user *buffer, size_t *length, loff_t *ppos)
5281{
 
5282	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
 
 
 
5283	int rc;
5284
5285	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5286	if (rc)
5287		return rc;
5288
5289	for_each_zone(zone)
5290		zone->min_unmapped_pages = (zone->present_pages *
5291				sysctl_min_unmapped_ratio) / 100;
5292	return 0;
5293}
5294
5295int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5296	void __user *buffer, size_t *length, loff_t *ppos)
5297{
 
5298	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
 
 
5299	int rc;
5300
5301	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5302	if (rc)
5303		return rc;
5304
5305	for_each_zone(zone)
5306		zone->min_slab_pages = (zone->present_pages *
5307				sysctl_min_slab_ratio) / 100;
5308	return 0;
5309}
5310#endif
5311
5312/*
5313 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5314 *	proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5315 *	whenever sysctl_lowmem_reserve_ratio changes.
5316 *
5317 * The reserve ratio obviously has absolutely no relation with the
5318 * minimum watermarks. The lowmem reserve ratio can only make sense
5319 * if in function of the boot time zone sizes.
5320 */
5321int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5322	void __user *buffer, size_t *length, loff_t *ppos)
5323{
 
 
5324	proc_dointvec_minmax(table, write, buffer, length, ppos);
 
 
 
 
 
 
5325	setup_per_zone_lowmem_reserve();
5326	return 0;
5327}
5328
5329/*
5330 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5331 * cpu.  It is the fraction of total pages in each zone that a hot per cpu pagelist
5332 * can have before it gets flushed back to buddy allocator.
5333 */
5334
5335int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5336	void __user *buffer, size_t *length, loff_t *ppos)
5337{
5338	struct zone *zone;
5339	unsigned int cpu;
5340	int ret;
5341
 
 
 
5342	ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5343	if (!write || (ret == -EINVAL))
5344		return ret;
5345	for_each_populated_zone(zone) {
5346		for_each_possible_cpu(cpu) {
5347			unsigned long  high;
5348			high = zone->present_pages / percpu_pagelist_fraction;
5349			setup_pagelist_highmark(
5350				per_cpu_ptr(zone->pageset, cpu), high);
5351		}
5352	}
5353	return 0;
5354}
5355
5356int hashdist = HASHDIST_DEFAULT;
 
 
5357
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5358#ifdef CONFIG_NUMA
5359static int __init set_hashdist(char *str)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5360{
5361	if (!str)
5362		return 0;
5363	hashdist = simple_strtoul(str, &str, 0);
5364	return 1;
5365}
5366__setup("hashdist=", set_hashdist);
5367#endif
5368
5369/*
5370 * allocate a large system hash table from bootmem
5371 * - it is assumed that the hash table must contain an exact power-of-2
5372 *   quantity of entries
5373 * - limit is the number of hash buckets, not the total allocation size
5374 */
5375void *__init alloc_large_system_hash(const char *tablename,
5376				     unsigned long bucketsize,
5377				     unsigned long numentries,
5378				     int scale,
5379				     int flags,
5380				     unsigned int *_hash_shift,
5381				     unsigned int *_hash_mask,
5382				     unsigned long limit)
5383{
5384	unsigned long long max = limit;
5385	unsigned long log2qty, size;
5386	void *table = NULL;
5387
5388	/* allow the kernel cmdline to have a say */
5389	if (!numentries) {
5390		/* round applicable memory size up to nearest megabyte */
5391		numentries = nr_kernel_pages;
5392		numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5393		numentries >>= 20 - PAGE_SHIFT;
5394		numentries <<= 20 - PAGE_SHIFT;
5395
5396		/* limit to 1 bucket per 2^scale bytes of low memory */
5397		if (scale > PAGE_SHIFT)
5398			numentries >>= (scale - PAGE_SHIFT);
5399		else
5400			numentries <<= (PAGE_SHIFT - scale);
5401
5402		/* Make sure we've got at least a 0-order allocation.. */
5403		if (unlikely(flags & HASH_SMALL)) {
5404			/* Makes no sense without HASH_EARLY */
5405			WARN_ON(!(flags & HASH_EARLY));
5406			if (!(numentries >> *_hash_shift)) {
5407				numentries = 1UL << *_hash_shift;
5408				BUG_ON(!numentries);
5409			}
5410		} else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5411			numentries = PAGE_SIZE / bucketsize;
5412	}
5413	numentries = roundup_pow_of_two(numentries);
5414
5415	/* limit allocation size to 1/16 total memory by default */
5416	if (max == 0) {
5417		max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5418		do_div(max, bucketsize);
5419	}
 
5420
5421	if (numentries > max)
5422		numentries = max;
5423
5424	log2qty = ilog2(numentries);
5425
5426	do {
5427		size = bucketsize << log2qty;
5428		if (flags & HASH_EARLY)
5429			table = alloc_bootmem_nopanic(size);
5430		else if (hashdist)
5431			table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5432		else {
5433			/*
5434			 * If bucketsize is not a power-of-two, we may free
5435			 * some pages at the end of hash table which
5436			 * alloc_pages_exact() automatically does
5437			 */
5438			if (get_order(size) < MAX_ORDER) {
5439				table = alloc_pages_exact(size, GFP_ATOMIC);
5440				kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5441			}
5442		}
5443	} while (!table && size > PAGE_SIZE && --log2qty);
5444
5445	if (!table)
5446		panic("Failed to allocate %s hash table\n", tablename);
 
 
 
 
 
 
 
 
 
5447
5448	printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5449	       tablename,
5450	       (1UL << log2qty),
5451	       ilog2(size) - PAGE_SHIFT,
5452	       size);
5453
5454	if (_hash_shift)
5455		*_hash_shift = log2qty;
5456	if (_hash_mask)
5457		*_hash_mask = (1 << log2qty) - 1;
5458
5459	return table;
5460}
5461
5462/* Return a pointer to the bitmap storing bits affecting a block of pages */
5463static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5464							unsigned long pfn)
5465{
5466#ifdef CONFIG_SPARSEMEM
5467	return __pfn_to_section(pfn)->pageblock_flags;
5468#else
5469	return zone->pageblock_flags;
5470#endif /* CONFIG_SPARSEMEM */
5471}
5472
5473static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5474{
5475#ifdef CONFIG_SPARSEMEM
5476	pfn &= (PAGES_PER_SECTION-1);
5477	return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5478#else
5479	pfn = pfn - zone->zone_start_pfn;
5480	return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5481#endif /* CONFIG_SPARSEMEM */
5482}
5483
5484/**
5485 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5486 * @page: The page within the block of interest
5487 * @start_bitidx: The first bit of interest to retrieve
5488 * @end_bitidx: The last bit of interest
5489 * returns pageblock_bits flags
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5490 */
5491unsigned long get_pageblock_flags_group(struct page *page,
5492					int start_bitidx, int end_bitidx)
5493{
5494	struct zone *zone;
5495	unsigned long *bitmap;
5496	unsigned long pfn, bitidx;
5497	unsigned long flags = 0;
5498	unsigned long value = 1;
5499
5500	zone = page_zone(page);
5501	pfn = page_to_pfn(page);
5502	bitmap = get_pageblock_bitmap(zone, pfn);
5503	bitidx = pfn_to_bitidx(zone, pfn);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5504
5505	for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5506		if (test_bit(bitidx + start_bitidx, bitmap))
5507			flags |= value;
 
 
5508
5509	return flags;
 
 
5510}
 
5511
5512/**
5513 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5514 * @page: The page within the block of interest
5515 * @start_bitidx: The first bit of interest
5516 * @end_bitidx: The last bit of interest
5517 * @flags: The flags to set
5518 */
5519void set_pageblock_flags_group(struct page *page, unsigned long flags,
5520					int start_bitidx, int end_bitidx)
5521{
5522	struct zone *zone;
5523	unsigned long *bitmap;
5524	unsigned long pfn, bitidx;
5525	unsigned long value = 1;
5526
5527	zone = page_zone(page);
5528	pfn = page_to_pfn(page);
5529	bitmap = get_pageblock_bitmap(zone, pfn);
5530	bitidx = pfn_to_bitidx(zone, pfn);
5531	VM_BUG_ON(pfn < zone->zone_start_pfn);
5532	VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5533
5534	for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5535		if (flags & value)
5536			__set_bit(bitidx + start_bitidx, bitmap);
5537		else
5538			__clear_bit(bitidx + start_bitidx, bitmap);
5539}
5540
5541/*
5542 * This is designed as sub function...plz see page_isolation.c also.
5543 * set/clear page block's type to be ISOLATE.
5544 * page allocater never alloc memory from ISOLATE block.
5545 */
5546
5547static int
5548__count_immobile_pages(struct zone *zone, struct page *page, int count)
5549{
5550	unsigned long pfn, iter, found;
5551	/*
5552	 * For avoiding noise data, lru_add_drain_all() should be called
5553	 * If ZONE_MOVABLE, the zone never contains immobile pages
5554	 */
5555	if (zone_idx(zone) == ZONE_MOVABLE)
5556		return true;
5557
5558	if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5559		return true;
 
 
5560
5561	pfn = page_to_pfn(page);
5562	for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5563		unsigned long check = pfn + iter;
5564
5565		if (!pfn_valid_within(check))
5566			continue;
5567
5568		page = pfn_to_page(check);
5569		if (!page_count(page)) {
5570			if (PageBuddy(page))
5571				iter += (1 << page_order(page)) - 1;
5572			continue;
5573		}
5574		if (!PageLRU(page))
5575			found++;
5576		/*
5577		 * If there are RECLAIMABLE pages, we need to check it.
5578		 * But now, memory offline itself doesn't call shrink_slab()
5579		 * and it still to be fixed.
5580		 */
5581		/*
5582		 * If the page is not RAM, page_count()should be 0.
5583		 * we don't need more check. This is an _used_ not-movable page.
5584		 *
5585		 * The problematic thing here is PG_reserved pages. PG_reserved
5586		 * is set to both of a memory hole page and a _used_ kernel
5587		 * page at boot.
5588		 */
5589		if (found > count)
5590			return false;
5591	}
5592	return true;
5593}
5594
5595bool is_pageblock_removable_nolock(struct page *page)
 
5596{
5597	struct zone *zone = page_zone(page);
5598	return __count_immobile_pages(zone, page, 0);
 
5599}
5600
5601int set_migratetype_isolate(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5602{
 
 
5603	struct zone *zone;
5604	unsigned long flags, pfn;
5605	struct memory_isolate_notify arg;
5606	int notifier_ret;
5607	int ret = -EBUSY;
5608
5609	zone = page_zone(page);
 
 
 
5610
5611	spin_lock_irqsave(&zone->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5612
5613	pfn = page_to_pfn(page);
5614	arg.start_pfn = pfn;
5615	arg.nr_pages = pageblock_nr_pages;
5616	arg.pages_found = 0;
5617
5618	/*
5619	 * It may be possible to isolate a pageblock even if the
5620	 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5621	 * notifier chain is used by balloon drivers to return the
5622	 * number of pages in a range that are held by the balloon
5623	 * driver to shrink memory. If all the pages are accounted for
5624	 * by balloons, are free, or on the LRU, isolation can continue.
5625	 * Later, for example, when memory hotplug notifier runs, these
5626	 * pages reported as "can be isolated" should be isolated(freed)
5627	 * by the balloon driver through the memory notifier chain.
5628	 */
5629	notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5630	notifier_ret = notifier_to_errno(notifier_ret);
5631	if (notifier_ret)
5632		goto out;
5633	/*
5634	 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5635	 * We just check MOVABLE pages.
5636	 */
5637	if (__count_immobile_pages(zone, page, arg.pages_found))
5638		ret = 0;
5639
5640	/*
5641	 * immobile means "not-on-lru" paes. If immobile is larger than
5642	 * removable-by-driver pages reported by notifier, we'll fail.
5643	 */
5644
5645out:
5646	if (!ret) {
5647		set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5648		move_freepages_block(zone, page, MIGRATE_ISOLATE);
5649	}
 
 
 
5650
5651	spin_unlock_irqrestore(&zone->lock, flags);
5652	if (!ret)
5653		drain_all_pages();
5654	return ret;
 
 
 
 
 
 
 
 
 
5655}
5656
5657void unset_migratetype_isolate(struct page *page)
5658{
5659	struct zone *zone;
5660	unsigned long flags;
5661	zone = page_zone(page);
5662	spin_lock_irqsave(&zone->lock, flags);
5663	if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5664		goto out;
5665	set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5666	move_freepages_block(zone, page, MIGRATE_MOVABLE);
5667out:
5668	spin_unlock_irqrestore(&zone->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
5669}
5670
5671#ifdef CONFIG_MEMORY_HOTREMOVE
5672/*
5673 * All pages in the range must be isolated before calling this.
 
5674 */
5675void
5676__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5677{
 
5678	struct page *page;
5679	struct zone *zone;
5680	int order, i;
5681	unsigned long pfn;
5682	unsigned long flags;
5683	/* find the first valid pfn */
5684	for (pfn = start_pfn; pfn < end_pfn; pfn++)
5685		if (pfn_valid(pfn))
5686			break;
5687	if (pfn == end_pfn)
5688		return;
5689	zone = page_zone(pfn_to_page(pfn));
5690	spin_lock_irqsave(&zone->lock, flags);
5691	pfn = start_pfn;
5692	while (pfn < end_pfn) {
5693		if (!pfn_valid(pfn)) {
 
 
 
 
 
5694			pfn++;
5695			continue;
5696		}
5697		page = pfn_to_page(pfn);
 
 
 
 
 
 
 
 
 
 
5698		BUG_ON(page_count(page));
5699		BUG_ON(!PageBuddy(page));
5700		order = page_order(page);
5701#ifdef CONFIG_DEBUG_VM
5702		printk(KERN_INFO "remove from free list %lx %d %lx\n",
5703		       pfn, 1 << order, end_pfn);
5704#endif
5705		list_del(&page->lru);
5706		rmv_page_order(page);
5707		zone->free_area[order].nr_free--;
5708		__mod_zone_page_state(zone, NR_FREE_PAGES,
5709				      - (1UL << order));
5710		for (i = 0; i < (1 << order); i++)
5711			SetPageReserved((page+i));
5712		pfn += (1 << order);
5713	}
5714	spin_unlock_irqrestore(&zone->lock, flags);
5715}
5716#endif
5717
5718#ifdef CONFIG_MEMORY_FAILURE
 
 
5719bool is_free_buddy_page(struct page *page)
5720{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5721	struct zone *zone = page_zone(page);
5722	unsigned long pfn = page_to_pfn(page);
5723	unsigned long flags;
5724	int order;
 
5725
5726	spin_lock_irqsave(&zone->lock, flags);
5727	for (order = 0; order < MAX_ORDER; order++) {
5728		struct page *page_head = page - (pfn & ((1 << order) - 1));
 
5729
5730		if (PageBuddy(page_head) && page_order(page_head) >= order)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5731			break;
5732	}
5733	spin_unlock_irqrestore(&zone->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5734
5735	return order < MAX_ORDER;
5736}
5737#endif
5738
5739static struct trace_print_flags pageflag_names[] = {
5740	{1UL << PG_locked,		"locked"	},
5741	{1UL << PG_error,		"error"		},
5742	{1UL << PG_referenced,		"referenced"	},
5743	{1UL << PG_uptodate,		"uptodate"	},
5744	{1UL << PG_dirty,		"dirty"		},
5745	{1UL << PG_lru,			"lru"		},
5746	{1UL << PG_active,		"active"	},
5747	{1UL << PG_slab,		"slab"		},
5748	{1UL << PG_owner_priv_1,	"owner_priv_1"	},
5749	{1UL << PG_arch_1,		"arch_1"	},
5750	{1UL << PG_reserved,		"reserved"	},
5751	{1UL << PG_private,		"private"	},
5752	{1UL << PG_private_2,		"private_2"	},
5753	{1UL << PG_writeback,		"writeback"	},
5754#ifdef CONFIG_PAGEFLAGS_EXTENDED
5755	{1UL << PG_head,		"head"		},
5756	{1UL << PG_tail,		"tail"		},
5757#else
5758	{1UL << PG_compound,		"compound"	},
5759#endif
5760	{1UL << PG_swapcache,		"swapcache"	},
5761	{1UL << PG_mappedtodisk,	"mappedtodisk"	},
5762	{1UL << PG_reclaim,		"reclaim"	},
5763	{1UL << PG_swapbacked,		"swapbacked"	},
5764	{1UL << PG_unevictable,		"unevictable"	},
5765#ifdef CONFIG_MMU
5766	{1UL << PG_mlocked,		"mlocked"	},
5767#endif
5768#ifdef CONFIG_ARCH_USES_PG_UNCACHED
5769	{1UL << PG_uncached,		"uncached"	},
5770#endif
5771#ifdef CONFIG_MEMORY_FAILURE
5772	{1UL << PG_hwpoison,		"hwpoison"	},
5773#endif
5774	{-1UL,				NULL		},
5775};
5776
5777static void dump_page_flags(unsigned long flags)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5778{
5779	const char *delim = "";
5780	unsigned long mask;
5781	int i;
 
 
 
 
 
 
 
 
5782
5783	printk(KERN_ALERT "page flags: %#lx(", flags);
 
 
 
5784
5785	/* remove zone id */
5786	flags &= (1UL << NR_PAGEFLAGS) - 1;
5787
5788	for (i = 0; pageflag_names[i].name && flags; i++) {
 
 
5789
5790		mask = pageflag_names[i].mask;
5791		if ((flags & mask) != mask)
5792			continue;
 
 
 
 
 
 
 
 
5793
5794		flags &= ~mask;
5795		printk("%s%s", delim, pageflag_names[i].name);
5796		delim = "|";
 
 
 
5797	}
5798
5799	/* check for left over flags */
5800	if (flags)
5801		printk("%s%#lx", delim, flags);
 
 
 
5802
5803	printk(")\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5804}
5805
5806void dump_page(struct page *page)
5807{
5808	printk(KERN_ALERT
5809	       "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5810		page, atomic_read(&page->_count), page_mapcount(page),
5811		page->mapping, page->index);
5812	dump_page_flags(page->flags);
5813	mem_cgroup_print_bad_page(page);
5814}
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/page_alloc.c
   4 *
   5 *  Manages the free list, the system allocates free pages here.
   6 *  Note that kmalloc() lives in slab.c
   7 *
   8 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   9 *  Swap reorganised 29.12.95, Stephen Tweedie
  10 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
  11 *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
  12 *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
  13 *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  14 *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  15 *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
  16 */
  17
  18#include <linux/stddef.h>
  19#include <linux/mm.h>
  20#include <linux/highmem.h>
  21#include <linux/interrupt.h>
 
  22#include <linux/jiffies.h>
 
 
  23#include <linux/compiler.h>
  24#include <linux/kernel.h>
  25#include <linux/kasan.h>
  26#include <linux/kmsan.h>
  27#include <linux/module.h>
  28#include <linux/suspend.h>
 
 
 
  29#include <linux/ratelimit.h>
  30#include <linux/oom.h>
 
  31#include <linux/topology.h>
  32#include <linux/sysctl.h>
  33#include <linux/cpu.h>
  34#include <linux/cpuset.h>
  35#include <linux/memory_hotplug.h>
  36#include <linux/nodemask.h>
 
  37#include <linux/vmstat.h>
 
 
 
 
 
  38#include <linux/fault-inject.h>
 
 
 
 
 
  39#include <linux/compaction.h>
  40#include <trace/events/kmem.h>
  41#include <trace/events/oom.h>
 
  42#include <linux/prefetch.h>
  43#include <linux/mm_inline.h>
  44#include <linux/mmu_notifier.h>
  45#include <linux/migrate.h>
  46#include <linux/sched/mm.h>
  47#include <linux/page_owner.h>
  48#include <linux/page_table_check.h>
  49#include <linux/memcontrol.h>
  50#include <linux/ftrace.h>
  51#include <linux/lockdep.h>
  52#include <linux/psi.h>
  53#include <linux/khugepaged.h>
  54#include <linux/delayacct.h>
  55#include <linux/cacheinfo.h>
  56#include <asm/div64.h>
  57#include "internal.h"
  58#include "shuffle.h"
  59#include "page_reporting.h"
  60
  61/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
  62typedef int __bitwise fpi_t;
  63
  64/* No special request */
  65#define FPI_NONE		((__force fpi_t)0)
  66
  67/*
  68 * Skip free page reporting notification for the (possibly merged) page.
  69 * This does not hinder free page reporting from grabbing the page,
  70 * reporting it and marking it "reported" -  it only skips notifying
  71 * the free page reporting infrastructure about a newly freed page. For
  72 * example, used when temporarily pulling a page from a freelist and
  73 * putting it back unmodified.
  74 */
  75#define FPI_SKIP_REPORT_NOTIFY	((__force fpi_t)BIT(0))
  76
  77/*
  78 * Place the (possibly merged) page to the tail of the freelist. Will ignore
  79 * page shuffling (relevant code - e.g., memory onlining - is expected to
  80 * shuffle the whole zone).
  81 *
  82 * Note: No code should rely on this flag for correctness - it's purely
  83 *       to allow for optimizations when handing back either fresh pages
  84 *       (memory onlining) or untouched pages (page isolation, free page
  85 *       reporting).
  86 */
  87#define FPI_TO_TAIL		((__force fpi_t)BIT(1))
  88
  89/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
  90static DEFINE_MUTEX(pcp_batch_high_lock);
  91#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
  92
  93#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
  94/*
  95 * On SMP, spin_trylock is sufficient protection.
  96 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
  97 */
  98#define pcp_trylock_prepare(flags)	do { } while (0)
  99#define pcp_trylock_finish(flag)	do { } while (0)
 100#else
 101
 102/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
 103#define pcp_trylock_prepare(flags)	local_irq_save(flags)
 104#define pcp_trylock_finish(flags)	local_irq_restore(flags)
 105#endif
 106
 107/*
 108 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
 109 * a migration causing the wrong PCP to be locked and remote memory being
 110 * potentially allocated, pin the task to the CPU for the lookup+lock.
 111 * preempt_disable is used on !RT because it is faster than migrate_disable.
 112 * migrate_disable is used on RT because otherwise RT spinlock usage is
 113 * interfered with and a high priority task cannot preempt the allocator.
 114 */
 115#ifndef CONFIG_PREEMPT_RT
 116#define pcpu_task_pin()		preempt_disable()
 117#define pcpu_task_unpin()	preempt_enable()
 118#else
 119#define pcpu_task_pin()		migrate_disable()
 120#define pcpu_task_unpin()	migrate_enable()
 121#endif
 122
 123/*
 124 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
 125 * Return value should be used with equivalent unlock helper.
 126 */
 127#define pcpu_spin_lock(type, member, ptr)				\
 128({									\
 129	type *_ret;							\
 130	pcpu_task_pin();						\
 131	_ret = this_cpu_ptr(ptr);					\
 132	spin_lock(&_ret->member);					\
 133	_ret;								\
 134})
 135
 136#define pcpu_spin_trylock(type, member, ptr)				\
 137({									\
 138	type *_ret;							\
 139	pcpu_task_pin();						\
 140	_ret = this_cpu_ptr(ptr);					\
 141	if (!spin_trylock(&_ret->member)) {				\
 142		pcpu_task_unpin();					\
 143		_ret = NULL;						\
 144	}								\
 145	_ret;								\
 146})
 147
 148#define pcpu_spin_unlock(member, ptr)					\
 149({									\
 150	spin_unlock(&ptr->member);					\
 151	pcpu_task_unpin();						\
 152})
 153
 154/* struct per_cpu_pages specific helpers. */
 155#define pcp_spin_lock(ptr)						\
 156	pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
 157
 158#define pcp_spin_trylock(ptr)						\
 159	pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
 160
 161#define pcp_spin_unlock(ptr)						\
 162	pcpu_spin_unlock(lock, ptr)
 163
 164#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
 165DEFINE_PER_CPU(int, numa_node);
 166EXPORT_PER_CPU_SYMBOL(numa_node);
 167#endif
 168
 169DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
 170
 171#ifdef CONFIG_HAVE_MEMORYLESS_NODES
 172/*
 173 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
 174 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
 175 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
 176 * defined in <linux/topology.h>.
 177 */
 178DEFINE_PER_CPU(int, _numa_mem_);		/* Kernel "local memory" node */
 179EXPORT_PER_CPU_SYMBOL(_numa_mem_);
 180#endif
 181
 182static DEFINE_MUTEX(pcpu_drain_mutex);
 183
 184#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
 185volatile unsigned long latent_entropy __latent_entropy;
 186EXPORT_SYMBOL(latent_entropy);
 187#endif
 188
 189/*
 190 * Array of node states.
 191 */
 192nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
 193	[N_POSSIBLE] = NODE_MASK_ALL,
 194	[N_ONLINE] = { { [0] = 1UL } },
 195#ifndef CONFIG_NUMA
 196	[N_NORMAL_MEMORY] = { { [0] = 1UL } },
 197#ifdef CONFIG_HIGHMEM
 198	[N_HIGH_MEMORY] = { { [0] = 1UL } },
 199#endif
 200	[N_MEMORY] = { { [0] = 1UL } },
 201	[N_CPU] = { { [0] = 1UL } },
 202#endif	/* NUMA */
 203};
 204EXPORT_SYMBOL(node_states);
 205
 
 
 
 206gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
 207
 
 208/*
 209 * A cached value of the page's pageblock's migratetype, used when the page is
 210 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
 211 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
 212 * Also the migratetype set in the page does not necessarily match the pcplist
 213 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
 214 * other index - this ensures that it will be put on the correct CMA freelist.
 215 */
 216static inline int get_pcppage_migratetype(struct page *page)
 
 
 
 217{
 218	return page->index;
 
 
 
 
 219}
 220
 221static inline void set_pcppage_migratetype(struct page *page, int migratetype)
 222{
 223	page->index = migratetype;
 
 
 
 224}
 
 225
 226#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 227unsigned int pageblock_order __read_mostly;
 228#endif
 229
 230static void __free_pages_ok(struct page *page, unsigned int order,
 231			    fpi_t fpi_flags);
 232
 233/*
 234 * results with 256, 32 in the lowmem_reserve sysctl:
 235 *	1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
 236 *	1G machine -> (16M dma, 784M normal, 224M high)
 237 *	NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
 238 *	HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
 239 *	HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
 240 *
 241 * TBD: should special case ZONE_DMA32 machines here - in those we normally
 242 * don't need any ZONE_NORMAL reservation
 243 */
 244static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
 245#ifdef CONFIG_ZONE_DMA
 246	[ZONE_DMA] = 256,
 247#endif
 248#ifdef CONFIG_ZONE_DMA32
 249	[ZONE_DMA32] = 256,
 250#endif
 251	[ZONE_NORMAL] = 32,
 252#ifdef CONFIG_HIGHMEM
 253	[ZONE_HIGHMEM] = 0,
 254#endif
 255	[ZONE_MOVABLE] = 0,
 256};
 257
 258char * const zone_names[MAX_NR_ZONES] = {
 
 
 259#ifdef CONFIG_ZONE_DMA
 260	 "DMA",
 261#endif
 262#ifdef CONFIG_ZONE_DMA32
 263	 "DMA32",
 264#endif
 265	 "Normal",
 266#ifdef CONFIG_HIGHMEM
 267	 "HighMem",
 268#endif
 269	 "Movable",
 270#ifdef CONFIG_ZONE_DEVICE
 271	 "Device",
 272#endif
 273};
 274
 275const char * const migratetype_names[MIGRATE_TYPES] = {
 276	"Unmovable",
 277	"Movable",
 278	"Reclaimable",
 279	"HighAtomic",
 280#ifdef CONFIG_CMA
 281	"CMA",
 282#endif
 283#ifdef CONFIG_MEMORY_ISOLATION
 284	"Isolate",
 285#endif
 286};
 287
 288int min_free_kbytes = 1024;
 289int user_min_free_kbytes = -1;
 290static int watermark_boost_factor __read_mostly = 15000;
 291static int watermark_scale_factor = 10;
 292
 293/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
 294int movable_zone;
 295EXPORT_SYMBOL(movable_zone);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 296
 297#if MAX_NUMNODES > 1
 298unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
 299unsigned int nr_online_nodes __read_mostly = 1;
 300EXPORT_SYMBOL(nr_node_ids);
 301EXPORT_SYMBOL(nr_online_nodes);
 302#endif
 303
 304static bool page_contains_unaccepted(struct page *page, unsigned int order);
 305static void accept_page(struct page *page, unsigned int order);
 306static bool try_to_accept_memory(struct zone *zone, unsigned int order);
 307static inline bool has_unaccepted_memory(void);
 308static bool __free_unaccepted(struct page *page);
 309
 310int page_group_by_mobility_disabled __read_mostly;
 311
 312#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 313/*
 314 * During boot we initialize deferred pages on-demand, as needed, but once
 315 * page_alloc_init_late() has finished, the deferred pages are all initialized,
 316 * and we can permanently disable that path.
 317 */
 318DEFINE_STATIC_KEY_TRUE(deferred_pages);
 319
 320static inline bool deferred_pages_enabled(void)
 321{
 322	return static_branch_unlikely(&deferred_pages);
 323}
 324
 325/*
 326 * deferred_grow_zone() is __init, but it is called from
 327 * get_page_from_freelist() during early boot until deferred_pages permanently
 328 * disables this call. This is why we have refdata wrapper to avoid warning,
 329 * and to ensure that the function body gets unloaded.
 330 */
 331static bool __ref
 332_deferred_grow_zone(struct zone *zone, unsigned int order)
 333{
 334       return deferred_grow_zone(zone, order);
 335}
 336#else
 337static inline bool deferred_pages_enabled(void)
 338{
 339	return false;
 340}
 341#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
 342
 343/* Return a pointer to the bitmap storing bits affecting a block of pages */
 344static inline unsigned long *get_pageblock_bitmap(const struct page *page,
 345							unsigned long pfn)
 346{
 347#ifdef CONFIG_SPARSEMEM
 348	return section_to_usemap(__pfn_to_section(pfn));
 349#else
 350	return page_zone(page)->pageblock_flags;
 351#endif /* CONFIG_SPARSEMEM */
 352}
 353
 354static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
 355{
 356#ifdef CONFIG_SPARSEMEM
 357	pfn &= (PAGES_PER_SECTION-1);
 358#else
 359	pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
 360#endif /* CONFIG_SPARSEMEM */
 361	return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
 362}
 363
 364/**
 365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
 366 * @page: The page within the block of interest
 367 * @pfn: The target page frame number
 368 * @mask: mask of bits that the caller is interested in
 369 *
 370 * Return: pageblock_bits flags
 371 */
 372unsigned long get_pfnblock_flags_mask(const struct page *page,
 373					unsigned long pfn, unsigned long mask)
 374{
 375	unsigned long *bitmap;
 376	unsigned long bitidx, word_bitidx;
 377	unsigned long word;
 378
 379	bitmap = get_pageblock_bitmap(page, pfn);
 380	bitidx = pfn_to_bitidx(page, pfn);
 381	word_bitidx = bitidx / BITS_PER_LONG;
 382	bitidx &= (BITS_PER_LONG-1);
 383	/*
 384	 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
 385	 * a consistent read of the memory array, so that results, even though
 386	 * racy, are not corrupted.
 387	 */
 388	word = READ_ONCE(bitmap[word_bitidx]);
 389	return (word >> bitidx) & mask;
 390}
 391
 392static __always_inline int get_pfnblock_migratetype(const struct page *page,
 393					unsigned long pfn)
 394{
 395	return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
 396}
 397
 398/**
 399 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
 400 * @page: The page within the block of interest
 401 * @flags: The flags to set
 402 * @pfn: The target page frame number
 403 * @mask: mask of bits that the caller is interested in
 404 */
 405void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
 406					unsigned long pfn,
 407					unsigned long mask)
 408{
 409	unsigned long *bitmap;
 410	unsigned long bitidx, word_bitidx;
 411	unsigned long word;
 412
 413	BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
 414	BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
 415
 416	bitmap = get_pageblock_bitmap(page, pfn);
 417	bitidx = pfn_to_bitidx(page, pfn);
 418	word_bitidx = bitidx / BITS_PER_LONG;
 419	bitidx &= (BITS_PER_LONG-1);
 420
 421	VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
 422
 423	mask <<= bitidx;
 424	flags <<= bitidx;
 425
 426	word = READ_ONCE(bitmap[word_bitidx]);
 427	do {
 428	} while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
 429}
 430
 431void set_pageblock_migratetype(struct page *page, int migratetype)
 432{
 433	if (unlikely(page_group_by_mobility_disabled &&
 434		     migratetype < MIGRATE_PCPTYPES))
 435		migratetype = MIGRATE_UNMOVABLE;
 436
 437	set_pfnblock_flags_mask(page, (unsigned long)migratetype,
 438				page_to_pfn(page), MIGRATETYPE_MASK);
 439}
 440
 441#ifdef CONFIG_DEBUG_VM
 442static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
 443{
 444	int ret;
 445	unsigned seq;
 446	unsigned long pfn = page_to_pfn(page);
 447	unsigned long sp, start_pfn;
 448
 449	do {
 450		seq = zone_span_seqbegin(zone);
 451		start_pfn = zone->zone_start_pfn;
 452		sp = zone->spanned_pages;
 453		ret = !zone_spans_pfn(zone, pfn);
 
 454	} while (zone_span_seqretry(zone, seq));
 455
 456	if (ret)
 457		pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
 458			pfn, zone_to_nid(zone), zone->name,
 459			start_pfn, start_pfn + sp);
 460
 461	return ret;
 462}
 463
 
 
 
 
 
 
 
 
 
 464/*
 465 * Temporary debugging check for pages not lying within a given zone.
 466 */
 467static int __maybe_unused bad_range(struct zone *zone, struct page *page)
 468{
 469	if (page_outside_zone_boundaries(zone, page))
 470		return 1;
 471	if (zone != page_zone(page))
 472		return 1;
 473
 474	return 0;
 475}
 476#else
 477static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
 478{
 479	return 0;
 480}
 481#endif
 482
 483static void bad_page(struct page *page, const char *reason)
 484{
 485	static unsigned long resume;
 486	static unsigned long nr_shown;
 487	static unsigned long nr_unshown;
 488
 
 
 
 
 
 
 489	/*
 490	 * Allow a burst of 60 reports, then keep quiet for that minute;
 491	 * or allow a steady drip of one report per second.
 492	 */
 493	if (nr_shown == 60) {
 494		if (time_before(jiffies, resume)) {
 495			nr_unshown++;
 496			goto out;
 497		}
 498		if (nr_unshown) {
 499			pr_alert(
 500			      "BUG: Bad page state: %lu messages suppressed\n",
 501				nr_unshown);
 502			nr_unshown = 0;
 503		}
 504		nr_shown = 0;
 505	}
 506	if (nr_shown++ == 0)
 507		resume = jiffies + 60 * HZ;
 508
 509	pr_alert("BUG: Bad page state in process %s  pfn:%05lx\n",
 510		current->comm, page_to_pfn(page));
 511	dump_page(page, reason);
 512
 513	print_modules();
 514	dump_stack();
 515out:
 516	/* Leave bad fields for debug, except PageBuddy could make trouble */
 517	page_mapcount_reset(page); /* remove PageBuddy */
 518	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 519}
 520
 521static inline unsigned int order_to_pindex(int migratetype, int order)
 522{
 523#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 524	if (order > PAGE_ALLOC_COSTLY_ORDER) {
 525		VM_BUG_ON(order != pageblock_order);
 526		return NR_LOWORDER_PCP_LISTS;
 527	}
 528#else
 529	VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
 530#endif
 531
 532	return (MIGRATE_PCPTYPES * order) + migratetype;
 533}
 534
 535static inline int pindex_to_order(unsigned int pindex)
 536{
 537	int order = pindex / MIGRATE_PCPTYPES;
 538
 539#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 540	if (pindex == NR_LOWORDER_PCP_LISTS)
 541		order = pageblock_order;
 542#else
 543	VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
 544#endif
 545
 546	return order;
 547}
 548
 549static inline bool pcp_allowed_order(unsigned int order)
 550{
 551	if (order <= PAGE_ALLOC_COSTLY_ORDER)
 552		return true;
 553#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 554	if (order == pageblock_order)
 555		return true;
 556#endif
 557	return false;
 558}
 559
 560static inline void free_the_page(struct page *page, unsigned int order)
 561{
 562	if (pcp_allowed_order(order))		/* Via pcp? */
 563		free_unref_page(page, order);
 564	else
 565		__free_pages_ok(page, order, FPI_NONE);
 566}
 567
 568/*
 569 * Higher-order pages are called "compound pages".  They are structured thusly:
 570 *
 571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
 572 *
 573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
 574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
 575 *
 576 * The first tail page's ->compound_order holds the order of allocation.
 
 
 
 
 577 * This usage means that zero-order pages may not be compound.
 578 */
 579
 580void prep_compound_page(struct page *page, unsigned int order)
 
 
 
 
 
 581{
 582	int i;
 583	int nr_pages = 1 << order;
 584
 
 
 585	__SetPageHead(page);
 586	for (i = 1; i < nr_pages; i++)
 587		prep_compound_tail(page, i);
 588
 589	prep_compound_head(page, order);
 
 
 590}
 591
 592void destroy_large_folio(struct folio *folio)
 
 593{
 594	if (folio_test_hugetlb(folio)) {
 595		free_huge_folio(folio);
 596		return;
 
 
 
 
 
 597	}
 598
 599	if (folio_test_large_rmappable(folio))
 600		folio_undo_large_rmappable(folio);
 601
 602	mem_cgroup_uncharge(folio);
 603	free_the_page(&folio->page, folio_order(folio));
 604}
 605
 606static inline void set_buddy_order(struct page *page, unsigned int order)
 607{
 608	set_page_private(page, order);
 609	__SetPageBuddy(page);
 610}
 611
 612#ifdef CONFIG_COMPACTION
 613static inline struct capture_control *task_capc(struct zone *zone)
 614{
 615	struct capture_control *capc = current->capture_control;
 616
 617	return unlikely(capc) &&
 618		!(current->flags & PF_KTHREAD) &&
 619		!capc->page &&
 620		capc->cc->zone == zone ? capc : NULL;
 621}
 622
 623static inline bool
 624compaction_capture(struct capture_control *capc, struct page *page,
 625		   int order, int migratetype)
 626{
 627	if (!capc || order != capc->cc->order)
 628		return false;
 629
 630	/* Do not accidentally pollute CMA or isolated regions*/
 631	if (is_migrate_cma(migratetype) ||
 632	    is_migrate_isolate(migratetype))
 633		return false;
 634
 635	/*
 636	 * Do not let lower order allocations pollute a movable pageblock.
 637	 * This might let an unmovable request use a reclaimable pageblock
 638	 * and vice-versa but no more than normal fallback logic which can
 639	 * have trouble finding a high-order free page.
 640	 */
 641	if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
 642		return false;
 643
 644	capc->page = page;
 645	return true;
 646}
 647
 648#else
 649static inline struct capture_control *task_capc(struct zone *zone)
 650{
 651	return NULL;
 
 652}
 653
 654static inline bool
 655compaction_capture(struct capture_control *capc, struct page *page,
 656		   int order, int migratetype)
 657{
 658	return false;
 659}
 660#endif /* CONFIG_COMPACTION */
 661
 662/* Used for pages not on another list */
 663static inline void add_to_free_list(struct page *page, struct zone *zone,
 664				    unsigned int order, int migratetype)
 665{
 666	struct free_area *area = &zone->free_area[order];
 667
 668	list_add(&page->buddy_list, &area->free_list[migratetype]);
 669	area->nr_free++;
 670}
 671
 672/* Used for pages not on another list */
 673static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
 674					 unsigned int order, int migratetype)
 675{
 676	struct free_area *area = &zone->free_area[order];
 677
 678	list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
 679	area->nr_free++;
 680}
 681
 682/*
 683 * Used for pages which are on another list. Move the pages to the tail
 684 * of the list - so the moved pages won't immediately be considered for
 685 * allocation again (e.g., optimization for memory onlining).
 
 
 
 
 
 
 
 
 
 
 
 
 686 */
 687static inline void move_to_free_list(struct page *page, struct zone *zone,
 688				     unsigned int order, int migratetype)
 689{
 690	struct free_area *area = &zone->free_area[order];
 691
 692	list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
 693}
 694
 695static inline void del_page_from_free_list(struct page *page, struct zone *zone,
 696					   unsigned int order)
 697{
 698	/* clear reported state and update reported page count */
 699	if (page_reported(page))
 700		__ClearPageReported(page);
 701
 702	list_del(&page->buddy_list);
 703	__ClearPageBuddy(page);
 704	set_page_private(page, 0);
 705	zone->free_area[order].nr_free--;
 706}
 707
 708static inline struct page *get_page_from_free_area(struct free_area *area,
 709					    int migratetype)
 710{
 711	return list_first_entry_or_null(&area->free_list[migratetype],
 712					struct page, buddy_list);
 713}
 714
 715/*
 716 * If this is not the largest possible page, check if the buddy
 717 * of the next-highest order is free. If it is, it's possible
 718 * that pages are being freed that will coalesce soon. In case,
 719 * that is happening, add the free page to the tail of the list
 720 * so it's less likely to be used soon and more likely to be merged
 721 * as a higher order page
 
 
 
 
 
 722 */
 723static inline bool
 724buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
 725		   struct page *page, unsigned int order)
 726{
 727	unsigned long higher_page_pfn;
 728	struct page *higher_page;
 729
 730	if (order >= MAX_PAGE_ORDER - 1)
 731		return false;
 732
 733	higher_page_pfn = buddy_pfn & pfn;
 734	higher_page = page + (higher_page_pfn - pfn);
 735
 736	return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
 737			NULL) != NULL;
 738}
 739
 740/*
 741 * Freeing function for a buddy system allocator.
 742 *
 743 * The concept of a buddy system is to maintain direct-mapped table
 744 * (containing bit values) for memory blocks of various "orders".
 745 * The bottom level table contains the map for the smallest allocatable
 746 * units of memory (here, pages), and each level above it describes
 747 * pairs of units from the levels below, hence, "buddies".
 748 * At a high level, all that happens here is marking the table entry
 749 * at the bottom level available, and propagating the changes upward
 750 * as necessary, plus some accounting needed to play nicely with other
 751 * parts of the VM system.
 752 * At each level, we keep a list of pages, which are heads of continuous
 753 * free pages of length of (1 << order) and marked with PageBuddy.
 754 * Page's order is recorded in page_private(page) field.
 755 * So when we are allocating or freeing one, we can derive the state of the
 756 * other.  That is, if we allocate a small block, and both were
 757 * free, the remainder of the region must be split into blocks.
 758 * If a block is freed, and its buddy is also free, then this
 759 * triggers coalescing into a block of larger size.
 760 *
 761 * -- nyc
 762 */
 763
 764static inline void __free_one_page(struct page *page,
 765		unsigned long pfn,
 766		struct zone *zone, unsigned int order,
 767		int migratetype, fpi_t fpi_flags)
 768{
 769	struct capture_control *capc = task_capc(zone);
 770	unsigned long buddy_pfn = 0;
 771	unsigned long combined_pfn;
 772	struct page *buddy;
 773	bool to_tail;
 774
 775	VM_BUG_ON(!zone_is_initialized(zone));
 776	VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
 
 777
 778	VM_BUG_ON(migratetype == -1);
 779	if (likely(!is_migrate_isolate(migratetype)))
 780		__mod_zone_freepage_state(zone, 1 << order, migratetype);
 781
 782	VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
 783	VM_BUG_ON_PAGE(bad_range(zone, page), page);
 784
 785	while (order < MAX_PAGE_ORDER) {
 786		if (compaction_capture(capc, page, order, migratetype)) {
 787			__mod_zone_freepage_state(zone, -(1 << order),
 788								migratetype);
 789			return;
 790		}
 791
 792		buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
 793		if (!buddy)
 794			goto done_merging;
 795
 796		if (unlikely(order >= pageblock_order)) {
 797			/*
 798			 * We want to prevent merge between freepages on pageblock
 799			 * without fallbacks and normal pageblock. Without this,
 800			 * pageblock isolation could cause incorrect freepage or CMA
 801			 * accounting or HIGHATOMIC accounting.
 802			 */
 803			int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
 804
 805			if (migratetype != buddy_mt
 806					&& (!migratetype_is_mergeable(migratetype) ||
 807						!migratetype_is_mergeable(buddy_mt)))
 808				goto done_merging;
 809		}
 810
 811		/*
 812		 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
 813		 * merge with it and move up one order.
 814		 */
 815		if (page_is_guard(buddy))
 816			clear_page_guard(zone, buddy, order, migratetype);
 817		else
 818			del_page_from_free_list(buddy, zone, order);
 819		combined_pfn = buddy_pfn & pfn;
 820		page = page + (combined_pfn - pfn);
 821		pfn = combined_pfn;
 822		order++;
 823	}
 
 824
 825done_merging:
 826	set_buddy_order(page, order);
 827
 828	if (fpi_flags & FPI_TO_TAIL)
 829		to_tail = true;
 830	else if (is_shuffle_order(order))
 831		to_tail = shuffle_pick_tail();
 832	else
 833		to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
 834
 835	if (to_tail)
 836		add_to_free_list_tail(page, zone, order, migratetype);
 837	else
 838		add_to_free_list(page, zone, order, migratetype);
 839
 840	/* Notify page reporting subsystem of freed page */
 841	if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
 842		page_reporting_notify_free(order);
 843}
 844
 845/**
 846 * split_free_page() -- split a free page at split_pfn_offset
 847 * @free_page:		the original free page
 848 * @order:		the order of the page
 849 * @split_pfn_offset:	split offset within the page
 850 *
 851 * Return -ENOENT if the free page is changed, otherwise 0
 852 *
 853 * It is used when the free page crosses two pageblocks with different migratetypes
 854 * at split_pfn_offset within the page. The split free page will be put into
 855 * separate migratetype lists afterwards. Otherwise, the function achieves
 856 * nothing.
 857 */
 858int split_free_page(struct page *free_page,
 859			unsigned int order, unsigned long split_pfn_offset)
 860{
 861	struct zone *zone = page_zone(free_page);
 862	unsigned long free_page_pfn = page_to_pfn(free_page);
 863	unsigned long pfn;
 864	unsigned long flags;
 865	int free_page_order;
 866	int mt;
 867	int ret = 0;
 868
 869	if (split_pfn_offset == 0)
 870		return ret;
 871
 872	spin_lock_irqsave(&zone->lock, flags);
 873
 874	if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
 875		ret = -ENOENT;
 876		goto out;
 877	}
 878
 879	mt = get_pfnblock_migratetype(free_page, free_page_pfn);
 880	if (likely(!is_migrate_isolate(mt)))
 881		__mod_zone_freepage_state(zone, -(1UL << order), mt);
 882
 883	del_page_from_free_list(free_page, zone, order);
 884	for (pfn = free_page_pfn;
 885	     pfn < free_page_pfn + (1UL << order);) {
 886		int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
 887
 888		free_page_order = min_t(unsigned int,
 889					pfn ? __ffs(pfn) : order,
 890					__fls(split_pfn_offset));
 891		__free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
 892				mt, FPI_NONE);
 893		pfn += 1UL << free_page_order;
 894		split_pfn_offset -= (1UL << free_page_order);
 895		/* we have done the first part, now switch to second part */
 896		if (split_pfn_offset == 0)
 897			split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
 898	}
 899out:
 900	spin_unlock_irqrestore(&zone->lock, flags);
 901	return ret;
 902}
 
 903/*
 904 * A bad page could be due to a number of fields. Instead of multiple branches,
 905 * try and check multiple fields with one check. The caller must do a detailed
 906 * check if necessary.
 907 */
 908static inline bool page_expected_state(struct page *page,
 909					unsigned long check_flags)
 910{
 911	if (unlikely(atomic_read(&page->_mapcount) != -1))
 912		return false;
 913
 914	if (unlikely((unsigned long)page->mapping |
 915			page_ref_count(page) |
 916#ifdef CONFIG_MEMCG
 917			page->memcg_data |
 918#endif
 919#ifdef CONFIG_PAGE_POOL
 920			((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
 921#endif
 922			(page->flags & check_flags)))
 923		return false;
 924
 925	return true;
 926}
 927
 928static const char *page_bad_reason(struct page *page, unsigned long flags)
 929{
 930	const char *bad_reason = NULL;
 931
 932	if (unlikely(atomic_read(&page->_mapcount) != -1))
 933		bad_reason = "nonzero mapcount";
 934	if (unlikely(page->mapping != NULL))
 935		bad_reason = "non-NULL mapping";
 936	if (unlikely(page_ref_count(page) != 0))
 937		bad_reason = "nonzero _refcount";
 938	if (unlikely(page->flags & flags)) {
 939		if (flags == PAGE_FLAGS_CHECK_AT_PREP)
 940			bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
 941		else
 942			bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
 943	}
 944#ifdef CONFIG_MEMCG
 945	if (unlikely(page->memcg_data))
 946		bad_reason = "page still charged to cgroup";
 947#endif
 948#ifdef CONFIG_PAGE_POOL
 949	if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
 950		bad_reason = "page_pool leak";
 951#endif
 952	return bad_reason;
 953}
 954
 955static void free_page_is_bad_report(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 956{
 957	bad_page(page,
 958		 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
 959}
 960
 961static inline bool free_page_is_bad(struct page *page)
 962{
 963	if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
 964		return false;
 965
 966	/* Something has gone sideways, find it */
 967	free_page_is_bad_report(page);
 968	return true;
 969}
 970
 971static inline bool is_check_pages_enabled(void)
 972{
 973	return static_branch_unlikely(&check_pages_enabled);
 974}
 
 
 
 
 
 
 
 
 
 975
 976static int free_tail_page_prepare(struct page *head_page, struct page *page)
 977{
 978	struct folio *folio = (struct folio *)head_page;
 979	int ret = 1;
 980
 981	/*
 982	 * We rely page->lru.next never has bit 0 set, unless the page
 983	 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
 984	 */
 985	BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
 986
 987	if (!is_check_pages_enabled()) {
 988		ret = 0;
 989		goto out;
 990	}
 991	switch (page - head_page) {
 992	case 1:
 993		/* the first tail page: these may be in place of ->mapping */
 994		if (unlikely(folio_entire_mapcount(folio))) {
 995			bad_page(page, "nonzero entire_mapcount");
 996			goto out;
 997		}
 998		if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
 999			bad_page(page, "nonzero nr_pages_mapped");
1000			goto out;
1001		}
1002		if (unlikely(atomic_read(&folio->_pincount))) {
1003			bad_page(page, "nonzero pincount");
1004			goto out;
1005		}
1006		break;
1007	case 2:
1008		/*
1009		 * the second tail page: ->mapping is
1010		 * deferred_list.next -- ignore value.
1011		 */
1012		break;
1013	default:
1014		if (page->mapping != TAIL_MAPPING) {
1015			bad_page(page, "corrupted mapping in tail page");
1016			goto out;
1017		}
1018		break;
1019	}
1020	if (unlikely(!PageTail(page))) {
1021		bad_page(page, "PageTail not set");
1022		goto out;
1023	}
1024	if (unlikely(compound_head(page) != head_page)) {
1025		bad_page(page, "compound_head not consistent");
1026		goto out;
1027	}
1028	ret = 0;
1029out:
1030	page->mapping = NULL;
1031	clear_compound_head(page);
1032	return ret;
1033}
1034
1035/*
1036 * Skip KASAN memory poisoning when either:
1037 *
1038 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1039 *    Tag-based KASAN modes skip pages freed via deferred memory initialization
1040 *    using page tags instead (see below).
1041 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1042 *    that error detection is disabled for accesses via the page address.
1043 *
1044 * Pages will have match-all tags in the following circumstances:
1045 *
1046 * 1. Pages are being initialized for the first time, including during deferred
1047 *    memory init; see the call to page_kasan_tag_reset in __init_single_page.
1048 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1049 *    exception of pages unpoisoned by kasan_unpoison_vmalloc.
1050 * 3. The allocation was excluded from being checked due to sampling,
1051 *    see the call to kasan_unpoison_pages.
1052 *
1053 * Poisoning pages during deferred memory init will greatly lengthen the
1054 * process and cause problem in large memory systems as the deferred pages
1055 * initialization is done with interrupt disabled.
1056 *
1057 * Assuming that there will be no reference to those newly initialized
1058 * pages before they are ever allocated, this should have no effect on
1059 * KASAN memory tracking as the poison will be properly inserted at page
1060 * allocation time. The only corner case is when pages are allocated by
1061 * on-demand allocation and then freed again before the deferred pages
1062 * initialization is done, but this is not likely to happen.
1063 */
1064static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1065{
1066	if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1067		return deferred_pages_enabled();
 
1068
1069	return page_kasan_tag(page) == KASAN_TAG_KERNEL;
 
 
1070}
1071
1072static void kernel_init_pages(struct page *page, int numpages)
1073{
1074	int i;
1075
1076	/* s390's use of memset() could override KASAN redzones. */
1077	kasan_disable_current();
1078	for (i = 0; i < numpages; i++)
1079		clear_highpage_kasan_tagged(page + i);
1080	kasan_enable_current();
1081}
1082
1083static __always_inline bool free_pages_prepare(struct page *page,
1084			unsigned int order, fpi_t fpi_flags)
1085{
1086	int bad = 0;
1087	bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1088	bool init = want_init_on_free();
1089	bool compound = PageCompound(page);
1090
1091	VM_BUG_ON_PAGE(PageTail(page), page);
1092
1093	trace_mm_page_free(page, order);
1094	kmsan_free_page(page, order);
1095
1096	if (memcg_kmem_online() && PageMemcgKmem(page))
1097		__memcg_kmem_uncharge_page(page, order);
1098
1099	if (unlikely(PageHWPoison(page)) && !order) {
1100		/* Do not let hwpoison pages hit pcplists/buddy */
1101		reset_page_owner(page, order);
1102		page_table_check_free(page, order);
1103		return false;
1104	}
1105
1106	VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1107
1108	/*
1109	 * Check tail pages before head page information is cleared to
1110	 * avoid checking PageCompound for order-0 pages.
1111	 */
1112	if (unlikely(order)) {
1113		int i;
1114
1115		if (compound)
1116			page[1].flags &= ~PAGE_FLAGS_SECOND;
1117		for (i = 1; i < (1 << order); i++) {
1118			if (compound)
1119				bad += free_tail_page_prepare(page, page + i);
1120			if (is_check_pages_enabled()) {
1121				if (free_page_is_bad(page + i)) {
1122					bad++;
1123					continue;
1124				}
1125			}
1126			(page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1127		}
1128	}
1129	if (PageMappingFlags(page))
1130		page->mapping = NULL;
1131	if (is_check_pages_enabled()) {
1132		if (free_page_is_bad(page))
1133			bad++;
1134		if (bad)
1135			return false;
1136	}
1137
1138	page_cpupid_reset_last(page);
1139	page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1140	reset_page_owner(page, order);
1141	page_table_check_free(page, order);
1142
1143	if (!PageHighMem(page)) {
1144		debug_check_no_locks_freed(page_address(page),
1145					   PAGE_SIZE << order);
1146		debug_check_no_obj_freed(page_address(page),
1147					   PAGE_SIZE << order);
1148	}
1149
1150	kernel_poison_pages(page, 1 << order);
1151
1152	/*
1153	 * As memory initialization might be integrated into KASAN,
1154	 * KASAN poisoning and memory initialization code must be
1155	 * kept together to avoid discrepancies in behavior.
1156	 *
1157	 * With hardware tag-based KASAN, memory tags must be set before the
1158	 * page becomes unavailable via debug_pagealloc or arch_free_page.
1159	 */
1160	if (!skip_kasan_poison) {
1161		kasan_poison_pages(page, order, init);
1162
1163		/* Memory is already initialized if KASAN did it internally. */
1164		if (kasan_has_integrated_init())
1165			init = false;
1166	}
1167	if (init)
1168		kernel_init_pages(page, 1 << order);
1169
1170	/*
1171	 * arch_free_page() can make the page's contents inaccessible.  s390
1172	 * does this.  So nothing which can access the page's contents should
1173	 * happen after this.
1174	 */
1175	arch_free_page(page, order);
1176
1177	debug_pagealloc_unmap_pages(page, 1 << order);
1178
1179	return true;
1180}
1181
1182/*
1183 * Frees a number of pages from the PCP lists
1184 * Assumes all pages on list are in same zone.
1185 * count is the number of pages to free.
1186 */
1187static void free_pcppages_bulk(struct zone *zone, int count,
1188					struct per_cpu_pages *pcp,
1189					int pindex)
1190{
1191	unsigned long flags;
1192	unsigned int order;
1193	bool isolated_pageblocks;
1194	struct page *page;
1195
1196	/*
1197	 * Ensure proper count is passed which otherwise would stuck in the
1198	 * below while (list_empty(list)) loop.
1199	 */
1200	count = min(pcp->count, count);
1201
1202	/* Ensure requested pindex is drained first. */
1203	pindex = pindex - 1;
1204
1205	spin_lock_irqsave(&zone->lock, flags);
1206	isolated_pageblocks = has_isolate_pageblock(zone);
1207
1208	while (count > 0) {
1209		struct list_head *list;
1210		int nr_pages;
1211
1212		/* Remove pages from lists in a round-robin fashion. */
1213		do {
1214			if (++pindex > NR_PCP_LISTS - 1)
1215				pindex = 0;
1216			list = &pcp->lists[pindex];
1217		} while (list_empty(list));
1218
1219		order = pindex_to_order(pindex);
1220		nr_pages = 1 << order;
1221		do {
1222			int mt;
1223
1224			page = list_last_entry(list, struct page, pcp_list);
1225			mt = get_pcppage_migratetype(page);
1226
1227			/* must delete to avoid corrupting pcp list */
1228			list_del(&page->pcp_list);
1229			count -= nr_pages;
1230			pcp->count -= nr_pages;
1231
1232			/* MIGRATE_ISOLATE page should not go to pcplists */
1233			VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1234			/* Pageblock could have been isolated meanwhile */
1235			if (unlikely(isolated_pageblocks))
1236				mt = get_pageblock_migratetype(page);
1237
1238			__free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1239			trace_mm_page_pcpu_drain(page, order, mt);
1240		} while (count > 0 && !list_empty(list));
1241	}
1242
1243	spin_unlock_irqrestore(&zone->lock, flags);
1244}
1245
1246static void free_one_page(struct zone *zone,
1247				struct page *page, unsigned long pfn,
1248				unsigned int order,
1249				int migratetype, fpi_t fpi_flags)
1250{
1251	unsigned long flags;
1252
1253	spin_lock_irqsave(&zone->lock, flags);
1254	if (unlikely(has_isolate_pageblock(zone) ||
1255		is_migrate_isolate(migratetype))) {
1256		migratetype = get_pfnblock_migratetype(page, pfn);
1257	}
1258	__free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1259	spin_unlock_irqrestore(&zone->lock, flags);
1260}
1261
1262static void __free_pages_ok(struct page *page, unsigned int order,
1263			    fpi_t fpi_flags)
1264{
1265	int migratetype;
1266	unsigned long pfn = page_to_pfn(page);
1267	struct zone *zone = page_zone(page);
1268
1269	if (!free_pages_prepare(page, order, fpi_flags))
1270		return;
1271
1272	/*
1273	 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1274	 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1275	 * This will reduce the lock holding time.
1276	 */
1277	migratetype = get_pfnblock_migratetype(page, pfn);
1278
1279	free_one_page(zone, page, pfn, order, migratetype, fpi_flags);
1280
1281	__count_vm_events(PGFREE, 1 << order);
 
 
 
1282}
1283
1284void __free_pages_core(struct page *page, unsigned int order)
 
 
 
1285{
1286	unsigned int nr_pages = 1 << order;
1287	struct page *p = page;
1288	unsigned int loop;
 
 
 
 
1289
1290	/*
1291	 * When initializing the memmap, __init_single_page() sets the refcount
1292	 * of all pages to 1 ("allocated"/"not free"). We have to set the
1293	 * refcount of all involved pages to 0.
1294	 */
1295	prefetchw(p);
1296	for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1297		prefetchw(p + 1);
1298		__ClearPageReserved(p);
1299		set_page_count(p, 0);
1300	}
1301	__ClearPageReserved(p);
1302	set_page_count(p, 0);
1303
1304	atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
 
 
 
 
1305
1306	if (page_contains_unaccepted(page, order)) {
1307		if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1308			return;
1309
1310		accept_page(page, order);
1311	}
1312
1313	/*
1314	 * Bypass PCP and place fresh pages right to the tail, primarily
1315	 * relevant for memory onlining.
1316	 */
1317	__free_pages_ok(page, order, FPI_TO_TAIL);
1318}
1319
1320/*
1321 * Check that the whole (or subset of) a pageblock given by the interval of
1322 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1323 * with the migration of free compaction scanner.
1324 *
1325 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1326 *
1327 * It's possible on some configurations to have a setup like node0 node1 node0
1328 * i.e. it's possible that all pages within a zones range of pages do not
1329 * belong to a single zone. We assume that a border between node0 and node1
1330 * can occur within a single pageblock, but not a node0 node1 node0
1331 * interleaving within a single pageblock. It is therefore sufficient to check
1332 * the first and last page of a pageblock and avoid checking each individual
1333 * page in a pageblock.
1334 *
1335 * Note: the function may return non-NULL struct page even for a page block
1336 * which contains a memory hole (i.e. there is no physical memory for a subset
1337 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1338 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1339 * even though the start pfn is online and valid. This should be safe most of
1340 * the time because struct pages are still initialized via init_unavailable_range()
1341 * and pfn walkers shouldn't touch any physical memory range for which they do
1342 * not recognize any specific metadata in struct pages.
1343 */
1344struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1345				     unsigned long end_pfn, struct zone *zone)
1346{
1347	struct page *start_page;
1348	struct page *end_page;
1349
1350	/* end_pfn is one past the range we are checking */
1351	end_pfn--;
1352
1353	if (!pfn_valid(end_pfn))
1354		return NULL;
1355
1356	start_page = pfn_to_online_page(start_pfn);
1357	if (!start_page)
1358		return NULL;
1359
1360	if (page_zone(start_page) != zone)
1361		return NULL;
1362
1363	end_page = pfn_to_page(end_pfn);
1364
1365	/* This gives a shorter code than deriving page_zone(end_page) */
1366	if (page_zone_id(start_page) != page_zone_id(end_page))
1367		return NULL;
1368
1369	return start_page;
1370}
1371
1372/*
1373 * The order of subdivision here is critical for the IO subsystem.
1374 * Please do not alter this order without good reasons and regression
1375 * testing. Specifically, as large blocks of memory are subdivided,
1376 * the order in which smaller blocks are delivered depends on the order
1377 * they're subdivided in this function. This is the primary factor
1378 * influencing the order in which pages are delivered to the IO
1379 * subsystem according to empirical testing, and this is also justified
1380 * by considering the behavior of a buddy system containing a single
1381 * large block of memory acted on by a series of small allocations.
1382 * This behavior is a critical factor in sglist merging's success.
1383 *
1384 * -- nyc
1385 */
1386static inline void expand(struct zone *zone, struct page *page,
1387	int low, int high, int migratetype)
 
1388{
1389	unsigned long size = 1 << high;
1390
1391	while (high > low) {
 
1392		high--;
1393		size >>= 1;
1394		VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1395
1396		/*
1397		 * Mark as guard pages (or page), that will allow to
1398		 * merge back to allocator when buddy will be freed.
1399		 * Corresponding page table entries will not be touched,
1400		 * pages will stay not present in virtual address space
1401		 */
1402		if (set_page_guard(zone, &page[size], high, migratetype))
1403			continue;
1404
1405		add_to_free_list(&page[size], zone, high, migratetype);
1406		set_buddy_order(&page[size], high);
1407	}
1408}
1409
1410static void check_new_page_bad(struct page *page)
1411{
1412	if (unlikely(page->flags & __PG_HWPOISON)) {
1413		/* Don't complain about hwpoisoned pages */
1414		page_mapcount_reset(page); /* remove PageBuddy */
1415		return;
1416	}
1417
1418	bad_page(page,
1419		 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1420}
1421
1422/*
1423 * This page is about to be returned from the page allocator
1424 */
1425static int check_new_page(struct page *page)
1426{
1427	if (likely(page_expected_state(page,
1428				PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1429		return 0;
1430
1431	check_new_page_bad(page);
1432	return 1;
 
 
 
1433}
1434
1435static inline bool check_new_pages(struct page *page, unsigned int order)
1436{
1437	if (is_check_pages_enabled()) {
1438		for (int i = 0; i < (1 << order); i++) {
1439			struct page *p = page + i;
1440
1441			if (check_new_page(p))
1442				return true;
1443		}
 
1444	}
1445
1446	return false;
1447}
1448
1449static inline bool should_skip_kasan_unpoison(gfp_t flags)
1450{
1451	/* Don't skip if a software KASAN mode is enabled. */
1452	if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1453	    IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1454		return false;
1455
1456	/* Skip, if hardware tag-based KASAN is not enabled. */
1457	if (!kasan_hw_tags_enabled())
1458		return true;
1459
1460	/*
1461	 * With hardware tag-based KASAN enabled, skip if this has been
1462	 * requested via __GFP_SKIP_KASAN.
1463	 */
1464	return flags & __GFP_SKIP_KASAN;
1465}
1466
1467static inline bool should_skip_init(gfp_t flags)
1468{
1469	/* Don't skip, if hardware tag-based KASAN is not enabled. */
1470	if (!kasan_hw_tags_enabled())
1471		return false;
1472
1473	/* For hardware tag-based KASAN, skip if requested. */
1474	return (flags & __GFP_SKIP_ZERO);
1475}
1476
1477inline void post_alloc_hook(struct page *page, unsigned int order,
1478				gfp_t gfp_flags)
1479{
1480	bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1481			!should_skip_init(gfp_flags);
1482	bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1483	int i;
1484
1485	set_page_private(page, 0);
1486	set_page_refcounted(page);
1487
1488	arch_alloc_page(page, order);
1489	debug_pagealloc_map_pages(page, 1 << order);
1490
1491	/*
1492	 * Page unpoisoning must happen before memory initialization.
1493	 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1494	 * allocations and the page unpoisoning code will complain.
1495	 */
1496	kernel_unpoison_pages(page, 1 << order);
1497
1498	/*
1499	 * As memory initialization might be integrated into KASAN,
1500	 * KASAN unpoisoning and memory initializion code must be
1501	 * kept together to avoid discrepancies in behavior.
1502	 */
1503
1504	/*
1505	 * If memory tags should be zeroed
1506	 * (which happens only when memory should be initialized as well).
1507	 */
1508	if (zero_tags) {
1509		/* Initialize both memory and memory tags. */
1510		for (i = 0; i != 1 << order; ++i)
1511			tag_clear_highpage(page + i);
1512
1513		/* Take note that memory was initialized by the loop above. */
1514		init = false;
1515	}
1516	if (!should_skip_kasan_unpoison(gfp_flags) &&
1517	    kasan_unpoison_pages(page, order, init)) {
1518		/* Take note that memory was initialized by KASAN. */
1519		if (kasan_has_integrated_init())
1520			init = false;
1521	} else {
1522		/*
1523		 * If memory tags have not been set by KASAN, reset the page
1524		 * tags to ensure page_address() dereferencing does not fault.
1525		 */
1526		for (i = 0; i != 1 << order; ++i)
1527			page_kasan_tag_reset(page + i);
1528	}
1529	/* If memory is still not initialized, initialize it now. */
1530	if (init)
1531		kernel_init_pages(page, 1 << order);
1532
1533	set_page_owner(page, order, gfp_flags);
1534	page_table_check_alloc(page, order);
1535}
1536
1537static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1538							unsigned int alloc_flags)
1539{
1540	post_alloc_hook(page, order, gfp_flags);
1541
1542	if (order && (gfp_flags & __GFP_COMP))
1543		prep_compound_page(page, order);
1544
1545	/*
1546	 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1547	 * allocate the page. The expectation is that the caller is taking
1548	 * steps that will free more memory. The caller should avoid the page
1549	 * being used for !PFMEMALLOC purposes.
1550	 */
1551	if (alloc_flags & ALLOC_NO_WATERMARKS)
1552		set_page_pfmemalloc(page);
1553	else
1554		clear_page_pfmemalloc(page);
1555}
1556
1557/*
1558 * Go through the free lists for the given migratetype and remove
1559 * the smallest available page from the freelists
1560 */
1561static __always_inline
1562struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1563						int migratetype)
1564{
1565	unsigned int current_order;
1566	struct free_area *area;
1567	struct page *page;
1568
1569	/* Find a page of the appropriate size in the preferred list */
1570	for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1571		area = &(zone->free_area[current_order]);
1572		page = get_page_from_free_area(area, migratetype);
1573		if (!page)
1574			continue;
1575		del_page_from_free_list(page, zone, current_order);
1576		expand(zone, page, order, current_order, migratetype);
1577		set_pcppage_migratetype(page, migratetype);
1578		trace_mm_page_alloc_zone_locked(page, order, migratetype,
1579				pcp_allowed_order(order) &&
1580				migratetype < MIGRATE_PCPTYPES);
 
1581		return page;
1582	}
1583
1584	return NULL;
1585}
1586
1587
1588/*
1589 * This array describes the order lists are fallen back to when
1590 * the free lists for the desirable migrate type are depleted
1591 *
1592 * The other migratetypes do not have fallbacks.
1593 */
1594static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1595	[MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE   },
1596	[MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1597	[MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE   },
 
1598};
1599
1600#ifdef CONFIG_CMA
1601static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1602					unsigned int order)
1603{
1604	return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1605}
1606#else
1607static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1608					unsigned int order) { return NULL; }
1609#endif
1610
1611/*
1612 * Move the free pages in a range to the freelist tail of the requested type.
1613 * Note that start_page and end_pages are not aligned on a pageblock
1614 * boundary. If alignment is required, use move_freepages_block()
1615 */
1616static int move_freepages(struct zone *zone,
1617			  unsigned long start_pfn, unsigned long end_pfn,
1618			  int migratetype, int *num_movable)
1619{
1620	struct page *page;
1621	unsigned long pfn;
1622	unsigned int order;
1623	int pages_moved = 0;
1624
1625	for (pfn = start_pfn; pfn <= end_pfn;) {
1626		page = pfn_to_page(pfn);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1627		if (!PageBuddy(page)) {
1628			/*
1629			 * We assume that pages that could be isolated for
1630			 * migration are movable. But we don't actually try
1631			 * isolating, as that would be expensive.
1632			 */
1633			if (num_movable &&
1634					(PageLRU(page) || __PageMovable(page)))
1635				(*num_movable)++;
1636			pfn++;
1637			continue;
1638		}
1639
1640		/* Make sure we are not inadvertently changing nodes */
1641		VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1642		VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1643
1644		order = buddy_order(page);
1645		move_to_free_list(page, zone, order, migratetype);
1646		pfn += 1 << order;
1647		pages_moved += 1 << order;
1648	}
1649
1650	return pages_moved;
1651}
1652
1653int move_freepages_block(struct zone *zone, struct page *page,
1654				int migratetype, int *num_movable)
1655{
1656	unsigned long start_pfn, end_pfn, pfn;
 
1657
1658	if (num_movable)
1659		*num_movable = 0;
1660
1661	pfn = page_to_pfn(page);
1662	start_pfn = pageblock_start_pfn(pfn);
1663	end_pfn = pageblock_end_pfn(pfn) - 1;
1664
1665	/* Do not cross zone boundaries */
1666	if (!zone_spans_pfn(zone, start_pfn))
1667		start_pfn = pfn;
1668	if (!zone_spans_pfn(zone, end_pfn))
1669		return 0;
1670
1671	return move_freepages(zone, start_pfn, end_pfn, migratetype,
1672								num_movable);
1673}
1674
1675static void change_pageblock_range(struct page *pageblock_page,
1676					int start_order, int migratetype)
1677{
1678	int nr_pageblocks = 1 << (start_order - pageblock_order);
1679
1680	while (nr_pageblocks--) {
1681		set_pageblock_migratetype(pageblock_page, migratetype);
1682		pageblock_page += pageblock_nr_pages;
1683	}
1684}
1685
1686/*
1687 * When we are falling back to another migratetype during allocation, try to
1688 * steal extra free pages from the same pageblocks to satisfy further
1689 * allocations, instead of polluting multiple pageblocks.
1690 *
1691 * If we are stealing a relatively large buddy page, it is likely there will
1692 * be more free pages in the pageblock, so try to steal them all. For
1693 * reclaimable and unmovable allocations, we steal regardless of page size,
1694 * as fragmentation caused by those allocations polluting movable pageblocks
1695 * is worse than movable allocations stealing from unmovable and reclaimable
1696 * pageblocks.
1697 */
1698static bool can_steal_fallback(unsigned int order, int start_mt)
1699{
1700	/*
1701	 * Leaving this order check is intended, although there is
1702	 * relaxed order check in next check. The reason is that
1703	 * we can actually steal whole pageblock if this condition met,
1704	 * but, below check doesn't guarantee it and that is just heuristic
1705	 * so could be changed anytime.
1706	 */
1707	if (order >= pageblock_order)
1708		return true;
1709
1710	if (order >= pageblock_order / 2 ||
1711		start_mt == MIGRATE_RECLAIMABLE ||
1712		start_mt == MIGRATE_UNMOVABLE ||
1713		page_group_by_mobility_disabled)
1714		return true;
1715
1716	return false;
1717}
1718
1719static inline bool boost_watermark(struct zone *zone)
1720{
1721	unsigned long max_boost;
1722
1723	if (!watermark_boost_factor)
1724		return false;
1725	/*
1726	 * Don't bother in zones that are unlikely to produce results.
1727	 * On small machines, including kdump capture kernels running
1728	 * in a small area, boosting the watermark can cause an out of
1729	 * memory situation immediately.
1730	 */
1731	if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1732		return false;
1733
1734	max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1735			watermark_boost_factor, 10000);
1736
1737	/*
1738	 * high watermark may be uninitialised if fragmentation occurs
1739	 * very early in boot so do not boost. We do not fall
1740	 * through and boost by pageblock_nr_pages as failing
1741	 * allocations that early means that reclaim is not going
1742	 * to help and it may even be impossible to reclaim the
1743	 * boosted watermark resulting in a hang.
1744	 */
1745	if (!max_boost)
1746		return false;
1747
1748	max_boost = max(pageblock_nr_pages, max_boost);
1749
1750	zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1751		max_boost);
1752
1753	return true;
1754}
1755
1756/*
1757 * This function implements actual steal behaviour. If order is large enough,
1758 * we can steal whole pageblock. If not, we first move freepages in this
1759 * pageblock to our migratetype and determine how many already-allocated pages
1760 * are there in the pageblock with a compatible migratetype. If at least half
1761 * of pages are free or compatible, we can change migratetype of the pageblock
1762 * itself, so pages freed in the future will be put on the correct free list.
1763 */
1764static void steal_suitable_fallback(struct zone *zone, struct page *page,
1765		unsigned int alloc_flags, int start_type, bool whole_block)
1766{
1767	unsigned int current_order = buddy_order(page);
1768	int free_pages, movable_pages, alike_pages;
1769	int old_block_type;
1770
1771	old_block_type = get_pageblock_migratetype(page);
1772
1773	/*
1774	 * This can happen due to races and we want to prevent broken
1775	 * highatomic accounting.
1776	 */
1777	if (is_migrate_highatomic(old_block_type))
1778		goto single_page;
1779
1780	/* Take ownership for orders >= pageblock_order */
1781	if (current_order >= pageblock_order) {
1782		change_pageblock_range(page, current_order, start_type);
1783		goto single_page;
1784	}
1785
1786	/*
1787	 * Boost watermarks to increase reclaim pressure to reduce the
1788	 * likelihood of future fallbacks. Wake kswapd now as the node
1789	 * may be balanced overall and kswapd will not wake naturally.
1790	 */
1791	if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1792		set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1793
1794	/* We are not allowed to try stealing from the whole block */
1795	if (!whole_block)
1796		goto single_page;
1797
1798	free_pages = move_freepages_block(zone, page, start_type,
1799						&movable_pages);
1800	/* moving whole block can fail due to zone boundary conditions */
1801	if (!free_pages)
1802		goto single_page;
1803
1804	/*
1805	 * Determine how many pages are compatible with our allocation.
1806	 * For movable allocation, it's the number of movable pages which
1807	 * we just obtained. For other types it's a bit more tricky.
1808	 */
1809	if (start_type == MIGRATE_MOVABLE) {
1810		alike_pages = movable_pages;
1811	} else {
1812		/*
1813		 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1814		 * to MOVABLE pageblock, consider all non-movable pages as
1815		 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1816		 * vice versa, be conservative since we can't distinguish the
1817		 * exact migratetype of non-movable pages.
1818		 */
1819		if (old_block_type == MIGRATE_MOVABLE)
1820			alike_pages = pageblock_nr_pages
1821						- (free_pages + movable_pages);
1822		else
1823			alike_pages = 0;
1824	}
1825	/*
1826	 * If a sufficient number of pages in the block are either free or of
1827	 * compatible migratability as our allocation, claim the whole block.
1828	 */
1829	if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1830			page_group_by_mobility_disabled)
1831		set_pageblock_migratetype(page, start_type);
1832
1833	return;
1834
1835single_page:
1836	move_to_free_list(page, zone, current_order, start_type);
1837}
1838
1839/*
1840 * Check whether there is a suitable fallback freepage with requested order.
1841 * If only_stealable is true, this function returns fallback_mt only if
1842 * we can steal other freepages all together. This would help to reduce
1843 * fragmentation due to mixed migratetype pages in one pageblock.
1844 */
1845int find_suitable_fallback(struct free_area *area, unsigned int order,
1846			int migratetype, bool only_stealable, bool *can_steal)
1847{
1848	int i;
1849	int fallback_mt;
1850
1851	if (area->nr_free == 0)
1852		return -1;
1853
1854	*can_steal = false;
1855	for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1856		fallback_mt = fallbacks[migratetype][i];
1857		if (free_area_empty(area, fallback_mt))
1858			continue;
1859
1860		if (can_steal_fallback(order, migratetype))
1861			*can_steal = true;
1862
1863		if (!only_stealable)
1864			return fallback_mt;
1865
1866		if (*can_steal)
1867			return fallback_mt;
1868	}
1869
1870	return -1;
1871}
1872
1873/*
1874 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1875 * there are no empty page blocks that contain a page with a suitable order
1876 */
1877static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1878{
1879	int mt;
1880	unsigned long max_managed, flags;
1881
1882	/*
1883	 * The number reserved as: minimum is 1 pageblock, maximum is
1884	 * roughly 1% of a zone. But if 1% of a zone falls below a
1885	 * pageblock size, then don't reserve any pageblocks.
1886	 * Check is race-prone but harmless.
1887	 */
1888	if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1889		return;
1890	max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1891	if (zone->nr_reserved_highatomic >= max_managed)
1892		return;
1893
1894	spin_lock_irqsave(&zone->lock, flags);
1895
1896	/* Recheck the nr_reserved_highatomic limit under the lock */
1897	if (zone->nr_reserved_highatomic >= max_managed)
1898		goto out_unlock;
1899
1900	/* Yoink! */
1901	mt = get_pageblock_migratetype(page);
1902	/* Only reserve normal pageblocks (i.e., they can merge with others) */
1903	if (migratetype_is_mergeable(mt)) {
1904		zone->nr_reserved_highatomic += pageblock_nr_pages;
1905		set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1906		move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1907	}
1908
1909out_unlock:
1910	spin_unlock_irqrestore(&zone->lock, flags);
1911}
1912
1913/*
1914 * Used when an allocation is about to fail under memory pressure. This
1915 * potentially hurts the reliability of high-order allocations when under
1916 * intense memory pressure but failed atomic allocations should be easier
1917 * to recover from than an OOM.
1918 *
1919 * If @force is true, try to unreserve a pageblock even though highatomic
1920 * pageblock is exhausted.
1921 */
1922static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1923						bool force)
1924{
1925	struct zonelist *zonelist = ac->zonelist;
1926	unsigned long flags;
1927	struct zoneref *z;
1928	struct zone *zone;
1929	struct page *page;
1930	int order;
1931	bool ret;
1932
1933	for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1934								ac->nodemask) {
1935		/*
1936		 * Preserve at least one pageblock unless memory pressure
1937		 * is really high.
1938		 */
1939		if (!force && zone->nr_reserved_highatomic <=
1940					pageblock_nr_pages)
1941			continue;
1942
1943		spin_lock_irqsave(&zone->lock, flags);
1944		for (order = 0; order < NR_PAGE_ORDERS; order++) {
1945			struct free_area *area = &(zone->free_area[order]);
1946
1947			page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1948			if (!page)
1949				continue;
1950
 
 
 
 
1951			/*
1952			 * In page freeing path, migratetype change is racy so
1953			 * we can counter several free pages in a pageblock
1954			 * in this loop although we changed the pageblock type
1955			 * from highatomic to ac->migratetype. So we should
1956			 * adjust the count once.
1957			 */
1958			if (is_migrate_highatomic_page(page)) {
1959				/*
1960				 * It should never happen but changes to
1961				 * locking could inadvertently allow a per-cpu
1962				 * drain to add pages to MIGRATE_HIGHATOMIC
1963				 * while unreserving so be safe and watch for
1964				 * underflows.
1965				 */
1966				zone->nr_reserved_highatomic -= min(
1967						pageblock_nr_pages,
1968						zone->nr_reserved_highatomic);
1969			}
1970
1971			/*
1972			 * Convert to ac->migratetype and avoid the normal
1973			 * pageblock stealing heuristics. Minimally, the caller
1974			 * is doing the work and needs the pages. More
1975			 * importantly, if the block was always converted to
1976			 * MIGRATE_UNMOVABLE or another type then the number
1977			 * of pageblocks that cannot be completely freed
1978			 * may increase.
1979			 */
1980			set_pageblock_migratetype(page, ac->migratetype);
1981			ret = move_freepages_block(zone, page, ac->migratetype,
1982									NULL);
1983			if (ret) {
1984				spin_unlock_irqrestore(&zone->lock, flags);
1985				return ret;
1986			}
1987		}
1988		spin_unlock_irqrestore(&zone->lock, flags);
1989	}
1990
1991	return false;
1992}
 
1993
1994/*
1995 * Try finding a free buddy page on the fallback list and put it on the free
1996 * list of requested migratetype, possibly along with other pages from the same
1997 * block, depending on fragmentation avoidance heuristics. Returns true if
1998 * fallback was found so that __rmqueue_smallest() can grab it.
1999 *
2000 * The use of signed ints for order and current_order is a deliberate
2001 * deviation from the rest of this file, to make the for loop
2002 * condition simpler.
2003 */
2004static __always_inline bool
2005__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2006						unsigned int alloc_flags)
2007{
2008	struct free_area *area;
2009	int current_order;
2010	int min_order = order;
2011	struct page *page;
2012	int fallback_mt;
2013	bool can_steal;
2014
2015	/*
2016	 * Do not steal pages from freelists belonging to other pageblocks
2017	 * i.e. orders < pageblock_order. If there are no local zones free,
2018	 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2019	 */
2020	if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2021		min_order = pageblock_order;
2022
2023	/*
2024	 * Find the largest available free page in the other list. This roughly
2025	 * approximates finding the pageblock with the most free pages, which
2026	 * would be too costly to do exactly.
2027	 */
2028	for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2029				--current_order) {
2030		area = &(zone->free_area[current_order]);
2031		fallback_mt = find_suitable_fallback(area, current_order,
2032				start_migratetype, false, &can_steal);
2033		if (fallback_mt == -1)
2034			continue;
2035
2036		/*
2037		 * We cannot steal all free pages from the pageblock and the
2038		 * requested migratetype is movable. In that case it's better to
2039		 * steal and split the smallest available page instead of the
2040		 * largest available page, because even if the next movable
2041		 * allocation falls back into a different pageblock than this
2042		 * one, it won't cause permanent fragmentation.
2043		 */
2044		if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2045					&& current_order > order)
2046			goto find_smallest;
2047
2048		goto do_steal;
2049	}
2050
2051	return false;
2052
2053find_smallest:
2054	for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2055		area = &(zone->free_area[current_order]);
2056		fallback_mt = find_suitable_fallback(area, current_order,
2057				start_migratetype, false, &can_steal);
2058		if (fallback_mt != -1)
2059			break;
2060	}
2061
2062	/*
2063	 * This should not happen - we already found a suitable fallback
2064	 * when looking for the largest page.
2065	 */
2066	VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2067
2068do_steal:
2069	page = get_page_from_free_area(area, fallback_mt);
2070
2071	steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2072								can_steal);
2073
2074	trace_mm_page_alloc_extfrag(page, order, current_order,
2075		start_migratetype, fallback_mt);
2076
2077	return true;
2078
2079}
2080
2081/*
2082 * Do the hard work of removing an element from the buddy allocator.
2083 * Call me with the zone->lock already held.
2084 */
2085static __always_inline struct page *
2086__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2087						unsigned int alloc_flags)
2088{
2089	struct page *page;
2090
2091	if (IS_ENABLED(CONFIG_CMA)) {
 
 
 
 
 
2092		/*
2093		 * Balance movable allocations between regular and CMA areas by
2094		 * allocating from CMA when over half of the zone's free memory
2095		 * is in the CMA area.
2096		 */
2097		if (alloc_flags & ALLOC_CMA &&
2098		    zone_page_state(zone, NR_FREE_CMA_PAGES) >
2099		    zone_page_state(zone, NR_FREE_PAGES) / 2) {
2100			page = __rmqueue_cma_fallback(zone, order);
2101			if (page)
2102				return page;
2103		}
2104	}
2105retry:
2106	page = __rmqueue_smallest(zone, order, migratetype);
2107	if (unlikely(!page)) {
2108		if (alloc_flags & ALLOC_CMA)
2109			page = __rmqueue_cma_fallback(zone, order);
2110
2111		if (!page && __rmqueue_fallback(zone, order, migratetype,
2112								alloc_flags))
2113			goto retry;
2114	}
2115	return page;
2116}
2117
2118/*
2119 * Obtain a specified number of elements from the buddy allocator, all under
2120 * a single hold of the lock, for efficiency.  Add them to the supplied list.
2121 * Returns the number of new pages which were placed at *list.
2122 */
2123static int rmqueue_bulk(struct zone *zone, unsigned int order,
2124			unsigned long count, struct list_head *list,
2125			int migratetype, unsigned int alloc_flags)
2126{
2127	unsigned long flags;
2128	int i;
2129
2130	spin_lock_irqsave(&zone->lock, flags);
2131	for (i = 0; i < count; ++i) {
2132		struct page *page = __rmqueue(zone, order, migratetype,
2133								alloc_flags);
2134		if (unlikely(page == NULL))
2135			break;
2136
2137		/*
2138		 * Split buddy pages returned by expand() are received here in
2139		 * physical page order. The page is added to the tail of
2140		 * caller's list. From the callers perspective, the linked list
2141		 * is ordered by page number under some conditions. This is
2142		 * useful for IO devices that can forward direction from the
2143		 * head, thus also in the physical page order. This is useful
2144		 * for IO devices that can merge IO requests if the physical
2145		 * pages are ordered properly.
2146		 */
2147		list_add_tail(&page->pcp_list, list);
2148		if (is_migrate_cma(get_pcppage_migratetype(page)))
2149			__mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2150					      -(1 << order));
 
 
2151	}
2152
2153	__mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2154	spin_unlock_irqrestore(&zone->lock, flags);
2155
2156	return i;
2157}
2158
2159/*
2160 * Called from the vmstat counter updater to decay the PCP high.
2161 * Return whether there are addition works to do.
2162 */
2163int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2164{
2165	int high_min, to_drain, batch;
2166	int todo = 0;
2167
2168	high_min = READ_ONCE(pcp->high_min);
2169	batch = READ_ONCE(pcp->batch);
2170	/*
2171	 * Decrease pcp->high periodically to try to free possible
2172	 * idle PCP pages.  And, avoid to free too many pages to
2173	 * control latency.  This caps pcp->high decrement too.
2174	 */
2175	if (pcp->high > high_min) {
2176		pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2177				 pcp->high - (pcp->high >> 3), high_min);
2178		if (pcp->high > high_min)
2179			todo++;
2180	}
2181
2182	to_drain = pcp->count - pcp->high;
2183	if (to_drain > 0) {
2184		spin_lock(&pcp->lock);
2185		free_pcppages_bulk(zone, to_drain, pcp, 0);
2186		spin_unlock(&pcp->lock);
2187		todo++;
2188	}
2189
2190	return todo;
2191}
2192
2193#ifdef CONFIG_NUMA
2194/*
2195 * Called from the vmstat counter updater to drain pagesets of this
2196 * currently executing processor on remote nodes after they have
2197 * expired.
 
 
 
2198 */
2199void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2200{
2201	int to_drain, batch;
 
2202
2203	batch = READ_ONCE(pcp->batch);
2204	to_drain = min(pcp->count, batch);
2205	if (to_drain > 0) {
2206		spin_lock(&pcp->lock);
2207		free_pcppages_bulk(zone, to_drain, pcp, 0);
2208		spin_unlock(&pcp->lock);
2209	}
 
2210}
2211#endif
2212
2213/*
2214 * Drain pcplists of the indicated processor and zone.
2215 */
2216static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2217{
2218	struct per_cpu_pages *pcp;
2219
2220	pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2221	if (pcp->count) {
2222		spin_lock(&pcp->lock);
2223		free_pcppages_bulk(zone, pcp->count, pcp, 0);
2224		spin_unlock(&pcp->lock);
2225	}
2226}
2227
2228/*
2229 * Drain pcplists of all zones on the indicated processor.
2230 */
2231static void drain_pages(unsigned int cpu)
2232{
 
2233	struct zone *zone;
2234
2235	for_each_populated_zone(zone) {
2236		drain_pages_zone(cpu, zone);
 
 
 
 
 
 
 
 
 
 
 
2237	}
2238}
2239
2240/*
2241 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2242 */
2243void drain_local_pages(struct zone *zone)
2244{
2245	int cpu = smp_processor_id();
2246
2247	if (zone)
2248		drain_pages_zone(cpu, zone);
2249	else
2250		drain_pages(cpu);
2251}
2252
2253/*
2254 * The implementation of drain_all_pages(), exposing an extra parameter to
2255 * drain on all cpus.
2256 *
2257 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2258 * not empty. The check for non-emptiness can however race with a free to
2259 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2260 * that need the guarantee that every CPU has drained can disable the
2261 * optimizing racy check.
2262 */
2263static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2264{
2265	int cpu;
2266
2267	/*
2268	 * Allocate in the BSS so we won't require allocation in
2269	 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2270	 */
2271	static cpumask_t cpus_with_pcps;
2272
2273	/*
2274	 * Do not drain if one is already in progress unless it's specific to
2275	 * a zone. Such callers are primarily CMA and memory hotplug and need
2276	 * the drain to be complete when the call returns.
2277	 */
2278	if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2279		if (!zone)
2280			return;
2281		mutex_lock(&pcpu_drain_mutex);
2282	}
2283
2284	/*
2285	 * We don't care about racing with CPU hotplug event
2286	 * as offline notification will cause the notified
2287	 * cpu to drain that CPU pcps and on_each_cpu_mask
2288	 * disables preemption as part of its processing
2289	 */
2290	for_each_online_cpu(cpu) {
2291		struct per_cpu_pages *pcp;
2292		struct zone *z;
2293		bool has_pcps = false;
2294
2295		if (force_all_cpus) {
2296			/*
2297			 * The pcp.count check is racy, some callers need a
2298			 * guarantee that no cpu is missed.
2299			 */
2300			has_pcps = true;
2301		} else if (zone) {
2302			pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2303			if (pcp->count)
2304				has_pcps = true;
2305		} else {
2306			for_each_populated_zone(z) {
2307				pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2308				if (pcp->count) {
2309					has_pcps = true;
2310					break;
2311				}
2312			}
2313		}
2314
2315		if (has_pcps)
2316			cpumask_set_cpu(cpu, &cpus_with_pcps);
2317		else
2318			cpumask_clear_cpu(cpu, &cpus_with_pcps);
2319	}
2320
2321	for_each_cpu(cpu, &cpus_with_pcps) {
2322		if (zone)
2323			drain_pages_zone(cpu, zone);
2324		else
2325			drain_pages(cpu);
2326	}
2327
2328	mutex_unlock(&pcpu_drain_mutex);
2329}
2330
2331/*
2332 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2333 *
2334 * When zone parameter is non-NULL, spill just the single zone's pages.
2335 */
2336void drain_all_pages(struct zone *zone)
2337{
2338	__drain_all_pages(zone, false);
2339}
2340
2341static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2342							unsigned int order)
2343{
2344	int migratetype;
2345
2346	if (!free_pages_prepare(page, order, FPI_NONE))
2347		return false;
2348
2349	migratetype = get_pfnblock_migratetype(page, pfn);
2350	set_pcppage_migratetype(page, migratetype);
2351	return true;
2352}
2353
2354static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2355{
2356	int min_nr_free, max_nr_free;
 
 
 
2357
2358	/* Free as much as possible if batch freeing high-order pages. */
2359	if (unlikely(free_high))
2360		return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2361
2362	/* Check for PCP disabled or boot pageset */
2363	if (unlikely(high < batch))
2364		return 1;
2365
2366	/* Leave at least pcp->batch pages on the list */
2367	min_nr_free = batch;
2368	max_nr_free = high - batch;
2369
2370	/*
2371	 * Increase the batch number to the number of the consecutive
2372	 * freed pages to reduce zone lock contention.
2373	 */
2374	batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2375
2376	return batch;
2377}
2378
2379static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2380		       int batch, bool free_high)
2381{
2382	int high, high_min, high_max;
2383
2384	high_min = READ_ONCE(pcp->high_min);
2385	high_max = READ_ONCE(pcp->high_max);
2386	high = pcp->high = clamp(pcp->high, high_min, high_max);
2387
2388	if (unlikely(!high))
2389		return 0;
2390
2391	if (unlikely(free_high)) {
2392		pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2393				high_min);
2394		return 0;
2395	}
2396
2397	/*
2398	 * If reclaim is active, limit the number of pages that can be
2399	 * stored on pcp lists
2400	 */
2401	if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2402		int free_count = max_t(int, pcp->free_count, batch);
2403
2404		pcp->high = max(high - free_count, high_min);
2405		return min(batch << 2, pcp->high);
2406	}
2407
2408	if (high_min == high_max)
2409		return high;
2410
2411	if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2412		int free_count = max_t(int, pcp->free_count, batch);
2413
2414		pcp->high = max(high - free_count, high_min);
2415		high = max(pcp->count, high_min);
2416	} else if (pcp->count >= high) {
2417		int need_high = pcp->free_count + batch;
2418
2419		/* pcp->high should be large enough to hold batch freed pages */
2420		if (pcp->high < need_high)
2421			pcp->high = clamp(need_high, high_min, high_max);
2422	}
2423
2424	return high;
2425}
2426
2427static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2428				   struct page *page, int migratetype,
2429				   unsigned int order)
2430{
2431	int high, batch;
2432	int pindex;
2433	bool free_high = false;
2434
2435	/*
2436	 * On freeing, reduce the number of pages that are batch allocated.
2437	 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2438	 * allocations.
2439	 */
2440	pcp->alloc_factor >>= 1;
2441	__count_vm_events(PGFREE, 1 << order);
2442	pindex = order_to_pindex(migratetype, order);
2443	list_add(&page->pcp_list, &pcp->lists[pindex]);
2444	pcp->count += 1 << order;
2445
2446	batch = READ_ONCE(pcp->batch);
2447	/*
2448	 * As high-order pages other than THP's stored on PCP can contribute
2449	 * to fragmentation, limit the number stored when PCP is heavily
2450	 * freeing without allocation. The remainder after bulk freeing
2451	 * stops will be drained from vmstat refresh context.
2452	 */
2453	if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2454		free_high = (pcp->free_count >= batch &&
2455			     (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2456			     (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2457			      pcp->count >= READ_ONCE(batch)));
2458		pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2459	} else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2460		pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2461	}
2462	if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2463		pcp->free_count += (1 << order);
2464	high = nr_pcp_high(pcp, zone, batch, free_high);
2465	if (pcp->count >= high) {
2466		free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2467				   pcp, pindex);
2468		if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2469		    zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2470				      ZONE_MOVABLE, 0))
2471			clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2472	}
 
2473}
 
2474
2475/*
2476 * Free a pcp page
 
2477 */
2478void free_unref_page(struct page *page, unsigned int order)
2479{
2480	unsigned long __maybe_unused UP_flags;
2481	struct per_cpu_pages *pcp;
2482	struct zone *zone;
2483	unsigned long pfn = page_to_pfn(page);
2484	int migratetype, pcpmigratetype;
2485
2486	if (!free_unref_page_prepare(page, pfn, order))
2487		return;
2488
 
 
 
 
 
 
 
2489	/*
2490	 * We only track unmovable, reclaimable and movable on pcp lists.
2491	 * Place ISOLATE pages on the isolated list because they are being
2492	 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2493	 * get those areas back if necessary. Otherwise, we may have to free
2494	 * excessively into the page allocator
2495	 */
2496	migratetype = pcpmigratetype = get_pcppage_migratetype(page);
2497	if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2498		if (unlikely(is_migrate_isolate(migratetype))) {
2499			free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2500			return;
2501		}
2502		pcpmigratetype = MIGRATE_MOVABLE;
2503	}
2504
2505	zone = page_zone(page);
2506	pcp_trylock_prepare(UP_flags);
2507	pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2508	if (pcp) {
2509		free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
2510		pcp_spin_unlock(pcp);
2511	} else {
2512		free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
 
2513	}
2514	pcp_trylock_finish(UP_flags);
2515}
2516
2517/*
2518 * Free a list of 0-order pages
2519 */
2520void free_unref_page_list(struct list_head *list)
2521{
2522	unsigned long __maybe_unused UP_flags;
2523	struct page *page, *next;
2524	struct per_cpu_pages *pcp = NULL;
2525	struct zone *locked_zone = NULL;
2526	int batch_count = 0;
2527	int migratetype;
2528
2529	/* Prepare pages for freeing */
2530	list_for_each_entry_safe(page, next, list, lru) {
2531		unsigned long pfn = page_to_pfn(page);
2532		if (!free_unref_page_prepare(page, pfn, 0)) {
2533			list_del(&page->lru);
2534			continue;
2535		}
2536
2537		/*
2538		 * Free isolated pages directly to the allocator, see
2539		 * comment in free_unref_page.
2540		 */
2541		migratetype = get_pcppage_migratetype(page);
2542		if (unlikely(is_migrate_isolate(migratetype))) {
2543			list_del(&page->lru);
2544			free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2545			continue;
2546		}
2547	}
2548
2549	list_for_each_entry_safe(page, next, list, lru) {
2550		struct zone *zone = page_zone(page);
2551
2552		list_del(&page->lru);
2553		migratetype = get_pcppage_migratetype(page);
2554
2555		/*
2556		 * Either different zone requiring a different pcp lock or
2557		 * excessive lock hold times when freeing a large list of
2558		 * pages.
2559		 */
2560		if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2561			if (pcp) {
2562				pcp_spin_unlock(pcp);
2563				pcp_trylock_finish(UP_flags);
2564			}
2565
2566			batch_count = 0;
2567
2568			/*
2569			 * trylock is necessary as pages may be getting freed
2570			 * from IRQ or SoftIRQ context after an IO completion.
2571			 */
2572			pcp_trylock_prepare(UP_flags);
2573			pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2574			if (unlikely(!pcp)) {
2575				pcp_trylock_finish(UP_flags);
2576				free_one_page(zone, page, page_to_pfn(page),
2577					      0, migratetype, FPI_NONE);
2578				locked_zone = NULL;
2579				continue;
2580			}
2581			locked_zone = zone;
2582		}
2583
2584		/*
2585		 * Non-isolated types over MIGRATE_PCPTYPES get added
2586		 * to the MIGRATE_MOVABLE pcp list.
2587		 */
2588		if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2589			migratetype = MIGRATE_MOVABLE;
2590
2591		trace_mm_page_free_batched(page);
2592		free_unref_page_commit(zone, pcp, page, migratetype, 0);
2593		batch_count++;
2594	}
2595
2596	if (pcp) {
2597		pcp_spin_unlock(pcp);
2598		pcp_trylock_finish(UP_flags);
2599	}
2600}
2601
2602/*
2603 * split_page takes a non-compound higher-order page, and splits it into
2604 * n (1<<order) sub-pages: page[0..n]
2605 * Each sub-page must be freed individually.
2606 *
2607 * Note: this is probably too low level an operation for use in drivers.
2608 * Please consult with lkml before using this in your driver.
2609 */
2610void split_page(struct page *page, unsigned int order)
2611{
2612	int i;
2613
2614	VM_BUG_ON_PAGE(PageCompound(page), page);
2615	VM_BUG_ON_PAGE(!page_count(page), page);
 
 
 
 
 
 
 
 
 
2616
2617	for (i = 1; i < (1 << order); i++)
2618		set_page_refcounted(page + i);
2619	split_page_owner(page, 1 << order);
2620	split_page_memcg(page, 1 << order);
2621}
2622EXPORT_SYMBOL_GPL(split_page);
2623
2624int __isolate_free_page(struct page *page, unsigned int order)
 
 
 
 
 
 
 
 
 
 
2625{
2626	struct zone *zone = page_zone(page);
2627	int mt = get_pageblock_migratetype(page);
 
 
 
 
 
 
2628
2629	if (!is_migrate_isolate(mt)) {
2630		unsigned long watermark;
2631		/*
2632		 * Obey watermarks as if the page was being allocated. We can
2633		 * emulate a high-order watermark check with a raised order-0
2634		 * watermark, because we already know our high-order page
2635		 * exists.
2636		 */
2637		watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2638		if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2639			return 0;
2640
2641		__mod_zone_freepage_state(zone, -(1UL << order), mt);
2642	}
 
 
 
2643
2644	del_page_from_free_list(page, zone, order);
 
 
2645
2646	/*
2647	 * Set the pageblock if the isolated page is at least half of a
2648	 * pageblock
2649	 */
2650	if (order >= pageblock_order - 1) {
2651		struct page *endpage = page + (1 << order) - 1;
2652		for (; page < endpage; page += pageblock_nr_pages) {
2653			int mt = get_pageblock_migratetype(page);
2654			/*
2655			 * Only change normal pageblocks (i.e., they can merge
2656			 * with others)
2657			 */
2658			if (migratetype_is_mergeable(mt))
2659				set_pageblock_migratetype(page,
2660							  MIGRATE_MOVABLE);
2661		}
2662	}
2663
2664	return 1UL << order;
2665}
2666
2667/**
2668 * __putback_isolated_page - Return a now-isolated page back where we got it
2669 * @page: Page that was isolated
2670 * @order: Order of the isolated page
2671 * @mt: The page's pageblock's migratetype
2672 *
2673 * This function is meant to return a page pulled from the free lists via
2674 * __isolate_free_page back to the free lists they were pulled from.
2675 */
2676void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2677{
2678	struct zone *zone = page_zone(page);
2679
2680	/* zone lock should be held when this function is called */
2681	lockdep_assert_held(&zone->lock);
2682
2683	/* Return isolated page to tail of freelist. */
2684	__free_one_page(page, page_to_pfn(page), zone, order, mt,
2685			FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2686}
2687
2688/*
2689 * Update NUMA hit/miss statistics
 
 
2690 */
2691static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2692				   long nr_account)
 
 
2693{
2694#ifdef CONFIG_NUMA
2695	enum numa_stat_item local_stat = NUMA_LOCAL;
 
2696
2697	/* skip numa counters update if numa stats is disabled */
2698	if (!static_branch_likely(&vm_numa_stat_key))
2699		return;
 
2700
2701	if (zone_to_nid(z) != numa_node_id())
2702		local_stat = NUMA_OTHER;
 
 
 
 
 
 
 
 
2703
2704	if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2705		__count_numa_events(z, NUMA_HIT, nr_account);
2706	else {
2707		__count_numa_events(z, NUMA_MISS, nr_account);
2708		__count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2709	}
2710	__count_numa_events(z, local_stat, nr_account);
2711#endif
2712}
2713
2714static __always_inline
2715struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2716			   unsigned int order, unsigned int alloc_flags,
2717			   int migratetype)
2718{
2719	struct page *page;
2720	unsigned long flags;
2721
2722	do {
2723		page = NULL;
2724		spin_lock_irqsave(&zone->lock, flags);
2725		if (alloc_flags & ALLOC_HIGHATOMIC)
2726			page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2727		if (!page) {
2728			page = __rmqueue(zone, order, migratetype, alloc_flags);
2729
 
 
 
 
2730			/*
2731			 * If the allocation fails, allow OOM handling access
2732			 * to HIGHATOMIC reserves as failing now is worse than
2733			 * failing a high-order atomic allocation in the
2734			 * future.
 
 
 
 
2735			 */
2736			if (!page && (alloc_flags & ALLOC_OOM))
2737				page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2738
2739			if (!page) {
2740				spin_unlock_irqrestore(&zone->lock, flags);
2741				return NULL;
2742			}
2743		}
2744		__mod_zone_freepage_state(zone, -(1 << order),
2745					  get_pcppage_migratetype(page));
2746		spin_unlock_irqrestore(&zone->lock, flags);
2747	} while (check_new_pages(page, order));
 
 
 
2748
2749	__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2750	zone_statistics(preferred_zone, zone, 1);
 
2751
 
 
 
2752	return page;
 
 
 
 
2753}
2754
2755static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2756{
2757	int high, base_batch, batch, max_nr_alloc;
2758	int high_max, high_min;
 
2759
2760	base_batch = READ_ONCE(pcp->batch);
2761	high_min = READ_ONCE(pcp->high_min);
2762	high_max = READ_ONCE(pcp->high_max);
2763	high = pcp->high = clamp(pcp->high, high_min, high_max);
 
 
 
 
 
 
2764
2765	/* Check for PCP disabled or boot pageset */
2766	if (unlikely(high < base_batch))
2767		return 1;
2768
2769	if (order)
2770		batch = base_batch;
2771	else
2772		batch = (base_batch << pcp->alloc_factor);
2773
2774	/*
2775	 * If we had larger pcp->high, we could avoid to allocate from
2776	 * zone.
2777	 */
2778	if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2779		high = pcp->high = min(high + batch, high_max);
 
 
 
 
 
 
 
 
 
 
 
 
 
2780
2781	if (!order) {
2782		max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2783		/*
2784		 * Double the number of pages allocated each time there is
2785		 * subsequent allocation of order-0 pages without any freeing.
2786		 */
2787		if (batch <= max_nr_alloc &&
2788		    pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2789			pcp->alloc_factor++;
2790		batch = min(batch, max_nr_alloc);
2791	}
2792
2793	/*
2794	 * Scale batch relative to order if batch implies free pages
2795	 * can be stored on the PCP. Batch can be 1 for small zones or
2796	 * for boot pagesets which should never store free pages as
2797	 * the pages may belong to arbitrary zones.
2798	 */
2799	if (batch > 1)
2800		batch = max(batch >> order, 2);
2801
2802	return batch;
2803}
2804
2805/* Remove page from the per-cpu list, caller must protect the list */
2806static inline
2807struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2808			int migratetype,
2809			unsigned int alloc_flags,
2810			struct per_cpu_pages *pcp,
2811			struct list_head *list)
2812{
2813	struct page *page;
2814
2815	do {
2816		if (list_empty(list)) {
2817			int batch = nr_pcp_alloc(pcp, zone, order);
2818			int alloced;
2819
2820			alloced = rmqueue_bulk(zone, order,
2821					batch, list,
2822					migratetype, alloc_flags);
2823
2824			pcp->count += alloced << order;
2825			if (unlikely(list_empty(list)))
2826				return NULL;
2827		}
2828
2829		page = list_first_entry(list, struct page, pcp_list);
2830		list_del(&page->pcp_list);
2831		pcp->count -= 1 << order;
2832	} while (check_new_pages(page, order));
2833
2834	return page;
2835}
2836
2837/* Lock and remove page from the per-cpu list */
2838static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2839			struct zone *zone, unsigned int order,
2840			int migratetype, unsigned int alloc_flags)
2841{
2842	struct per_cpu_pages *pcp;
2843	struct list_head *list;
2844	struct page *page;
2845	unsigned long __maybe_unused UP_flags;
2846
2847	/* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2848	pcp_trylock_prepare(UP_flags);
2849	pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2850	if (!pcp) {
2851		pcp_trylock_finish(UP_flags);
2852		return NULL;
2853	}
2854
2855	/*
2856	 * On allocation, reduce the number of pages that are batch freed.
2857	 * See nr_pcp_free() where free_factor is increased for subsequent
2858	 * frees.
2859	 */
2860	pcp->free_count >>= 1;
2861	list = &pcp->lists[order_to_pindex(migratetype, order)];
2862	page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2863	pcp_spin_unlock(pcp);
2864	pcp_trylock_finish(UP_flags);
2865	if (page) {
2866		__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2867		zone_statistics(preferred_zone, zone, 1);
2868	}
2869	return page;
2870}
2871
2872/*
2873 * Allocate a page from the given zone.
2874 * Use pcplists for THP or "cheap" high-order allocations.
2875 */
2876
2877/*
2878 * Do not instrument rmqueue() with KMSAN. This function may call
2879 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2880 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2881 * may call rmqueue() again, which will result in a deadlock.
2882 */
2883__no_sanitize_memory
2884static inline
2885struct page *rmqueue(struct zone *preferred_zone,
2886			struct zone *zone, unsigned int order,
2887			gfp_t gfp_flags, unsigned int alloc_flags,
2888			int migratetype)
2889{
2890	struct page *page;
 
 
2891
2892	/*
2893	 * We most definitely don't want callers attempting to
2894	 * allocate greater than order-1 page units with __GFP_NOFAIL.
2895	 */
2896	WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2897
2898	if (likely(pcp_allowed_order(order))) {
2899		page = rmqueue_pcplist(preferred_zone, zone, order,
2900				       migratetype, alloc_flags);
2901		if (likely(page))
2902			goto out;
2903	}
2904
2905	page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2906							migratetype);
2907
2908out:
2909	/* Separate test+clear to avoid unnecessary atomics */
2910	if ((alloc_flags & ALLOC_KSWAPD) &&
2911	    unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2912		clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2913		wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2914	}
2915
2916	VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2917	return page;
2918}
2919
2920noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
 
2921{
2922	return __should_fail_alloc_page(gfp_mask, order);
 
2923}
2924ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2925
2926static inline long __zone_watermark_unusable_free(struct zone *z,
2927				unsigned int order, unsigned int alloc_flags)
2928{
2929	long unusable_free = (1 << order) - 1;
2930
2931	/*
2932	 * If the caller does not have rights to reserves below the min
2933	 * watermark then subtract the high-atomic reserves. This will
2934	 * over-estimate the size of the atomic reserve but it avoids a search.
2935	 */
2936	if (likely(!(alloc_flags & ALLOC_RESERVES)))
2937		unusable_free += z->nr_reserved_highatomic;
2938
2939#ifdef CONFIG_CMA
2940	/* If allocation can't use CMA areas don't use free CMA pages */
2941	if (!(alloc_flags & ALLOC_CMA))
2942		unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2943#endif
2944#ifdef CONFIG_UNACCEPTED_MEMORY
2945	unusable_free += zone_page_state(z, NR_UNACCEPTED);
2946#endif
2947
2948	return unusable_free;
2949}
2950
 
2951/*
2952 * Return true if free base pages are above 'mark'. For high-order checks it
2953 * will return true of the order-0 watermark is reached and there is at least
2954 * one free page of a suitable size. Checking now avoids taking the zone lock
2955 * to check in the allocation paths if no pages are free.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2956 */
2957bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2958			 int highest_zoneidx, unsigned int alloc_flags,
2959			 long free_pages)
2960{
2961	long min = mark;
2962	int o;
2963
2964	/* free_pages may go negative - that's OK */
2965	free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
 
2966
2967	if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2968		/*
2969		 * __GFP_HIGH allows access to 50% of the min reserve as well
2970		 * as OOM.
2971		 */
2972		if (alloc_flags & ALLOC_MIN_RESERVE) {
2973			min -= min / 2;
 
 
 
2974
2975			/*
2976			 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2977			 * access more reserves than just __GFP_HIGH. Other
2978			 * non-blocking allocations requests such as GFP_NOWAIT
2979			 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2980			 * access to the min reserve.
2981			 */
2982			if (alloc_flags & ALLOC_NON_BLOCK)
2983				min -= min / 4;
2984		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2985
2986		/*
2987		 * OOM victims can try even harder than the normal reserve
2988		 * users on the grounds that it's definitely going to be in
2989		 * the exit path shortly and free memory. Any allocation it
2990		 * makes during the free path will be small and short-lived.
2991		 */
2992		if (alloc_flags & ALLOC_OOM)
2993			min -= min / 2;
2994	}
2995
2996	/*
2997	 * Check watermarks for an order-0 allocation request. If these
2998	 * are not met, then a high-order request also cannot go ahead
2999	 * even if a suitable page happened to be free.
3000	 */
3001	if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3002		return false;
3003
3004	/* If this is an order-0 request then the watermark is fine */
3005	if (!order)
3006		return true;
3007
3008	/* For a high-order request, check at least one suitable page is free */
3009	for (o = order; o < NR_PAGE_ORDERS; o++) {
3010		struct free_area *area = &z->free_area[o];
3011		int mt;
3012
3013		if (!area->nr_free)
3014			continue;
3015
3016		for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3017			if (!free_area_empty(area, mt))
3018				return true;
3019		}
3020
3021#ifdef CONFIG_CMA
3022		if ((alloc_flags & ALLOC_CMA) &&
3023		    !free_area_empty(area, MIGRATE_CMA)) {
3024			return true;
3025		}
3026#endif
3027		if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3028		    !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3029			return true;
3030		}
3031	}
3032	return false;
3033}
3034
3035bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3036		      int highest_zoneidx, unsigned int alloc_flags)
 
 
 
 
3037{
3038	return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3039					zone_page_state(z, NR_FREE_PAGES));
3040}
3041
3042static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3043				unsigned long mark, int highest_zoneidx,
3044				unsigned int alloc_flags, gfp_t gfp_mask)
3045{
3046	long free_pages;
3047
3048	free_pages = zone_page_state(z, NR_FREE_PAGES);
3049
3050	/*
3051	 * Fast check for order-0 only. If this fails then the reserves
3052	 * need to be calculated.
3053	 */
3054	if (!order) {
3055		long usable_free;
3056		long reserved;
3057
3058		usable_free = free_pages;
3059		reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3060
3061		/* reserved may over estimate high-atomic reserves. */
3062		usable_free -= min(usable_free, reserved);
3063		if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3064			return true;
3065	}
3066
3067	if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3068					free_pages))
3069		return true;
3070
3071	/*
3072	 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3073	 * when checking the min watermark. The min watermark is the
3074	 * point where boosting is ignored so that kswapd is woken up
3075	 * when below the low watermark.
3076	 */
3077	if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3078		&& ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3079		mark = z->_watermark[WMARK_MIN];
3080		return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3081					alloc_flags, free_pages);
3082	}
3083
3084	return false;
3085}
3086
3087bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3088			unsigned long mark, int highest_zoneidx)
 
 
 
3089{
3090	long free_pages = zone_page_state(z, NR_FREE_PAGES);
3091
3092	if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3093		free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
 
3094
3095	return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3096								free_pages);
3097}
3098
3099#ifdef CONFIG_NUMA
3100int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3101
3102static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3103{
3104	return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3105				node_reclaim_distance;
3106}
3107#else	/* CONFIG_NUMA */
3108static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
 
3109{
3110	return true;
3111}
3112#endif	/* CONFIG_NUMA */
3113
3114/*
3115 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3116 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3117 * premature use of a lower zone may cause lowmem pressure problems that
3118 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3119 * probably too small. It only makes sense to spread allocations to avoid
3120 * fragmentation between the Normal and DMA32 zones.
3121 */
3122static inline unsigned int
3123alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3124{
3125	unsigned int alloc_flags;
3126
3127	/*
3128	 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3129	 * to save a branch.
3130	 */
3131	alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3132
3133#ifdef CONFIG_ZONE_DMA32
3134	if (!zone)
3135		return alloc_flags;
3136
3137	if (zone_idx(zone) != ZONE_NORMAL)
3138		return alloc_flags;
3139
3140	/*
3141	 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3142	 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3143	 * on UMA that if Normal is populated then so is DMA32.
3144	 */
3145	BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3146	if (nr_online_nodes > 1 && !populated_zone(--zone))
3147		return alloc_flags;
3148
3149	alloc_flags |= ALLOC_NOFRAGMENT;
3150#endif /* CONFIG_ZONE_DMA32 */
3151	return alloc_flags;
3152}
3153
3154/* Must be called after current_gfp_context() which can change gfp_mask */
3155static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3156						  unsigned int alloc_flags)
3157{
3158#ifdef CONFIG_CMA
3159	if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3160		alloc_flags |= ALLOC_CMA;
3161#endif
3162	return alloc_flags;
3163}
 
3164
3165/*
3166 * get_page_from_freelist goes through the zonelist trying to allocate
3167 * a page.
3168 */
3169static struct page *
3170get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3171						const struct alloc_context *ac)
 
3172{
3173	struct zoneref *z;
 
 
3174	struct zone *zone;
3175	struct pglist_data *last_pgdat = NULL;
3176	bool last_pgdat_dirty_ok = false;
3177	bool no_fallback;
3178
3179retry:
 
3180	/*
3181	 * Scan zonelist, looking for a zone with enough free.
3182	 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3183	 */
3184	no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3185	z = ac->preferred_zoneref;
3186	for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3187					ac->nodemask) {
3188		struct page *page;
3189		unsigned long mark;
3190
3191		if (cpusets_enabled() &&
3192			(alloc_flags & ALLOC_CPUSET) &&
3193			!__cpuset_zone_allowed(zone, gfp_mask))
3194				continue;
3195		/*
3196		 * When allocating a page cache page for writing, we
3197		 * want to get it from a node that is within its dirty
3198		 * limit, such that no single node holds more than its
3199		 * proportional share of globally allowed dirty pages.
3200		 * The dirty limits take into account the node's
3201		 * lowmem reserves and high watermark so that kswapd
3202		 * should be able to balance it without having to
3203		 * write pages from its LRU list.
3204		 *
3205		 * XXX: For now, allow allocations to potentially
3206		 * exceed the per-node dirty limit in the slowpath
3207		 * (spread_dirty_pages unset) before going into reclaim,
3208		 * which is important when on a NUMA setup the allowed
3209		 * nodes are together not big enough to reach the
3210		 * global limit.  The proper fix for these situations
3211		 * will require awareness of nodes in the
3212		 * dirty-throttling and the flusher threads.
3213		 */
3214		if (ac->spread_dirty_pages) {
3215			if (last_pgdat != zone->zone_pgdat) {
3216				last_pgdat = zone->zone_pgdat;
3217				last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3218			}
3219
3220			if (!last_pgdat_dirty_ok)
3221				continue;
3222		}
 
3223
3224		if (no_fallback && nr_online_nodes > 1 &&
3225		    zone != ac->preferred_zoneref->zone) {
3226			int local_nid;
 
3227
3228			/*
3229			 * If moving to a remote node, retry but allow
3230			 * fragmenting fallbacks. Locality is more important
3231			 * than fragmentation avoidance.
3232			 */
3233			local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3234			if (zone_to_nid(zone) != local_nid) {
3235				alloc_flags &= ~ALLOC_NOFRAGMENT;
3236				goto retry;
3237			}
3238		}
3239
3240		/*
3241		 * Detect whether the number of free pages is below high
3242		 * watermark.  If so, we will decrease pcp->high and free
3243		 * PCP pages in free path to reduce the possibility of
3244		 * premature page reclaiming.  Detection is done here to
3245		 * avoid to do that in hotter free path.
3246		 */
3247		if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3248			goto check_alloc_wmark;
3249
3250		mark = high_wmark_pages(zone);
3251		if (zone_watermark_fast(zone, order, mark,
3252					ac->highest_zoneidx, alloc_flags,
3253					gfp_mask))
3254			goto try_this_zone;
3255		else
3256			set_bit(ZONE_BELOW_HIGH, &zone->flags);
3257
3258check_alloc_wmark:
3259		mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3260		if (!zone_watermark_fast(zone, order, mark,
3261				       ac->highest_zoneidx, alloc_flags,
3262				       gfp_mask)) {
3263			int ret;
3264
3265			if (has_unaccepted_memory()) {
3266				if (try_to_accept_memory(zone, order))
3267					goto try_this_zone;
3268			}
3269
3270#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3271			/*
3272			 * Watermark failed for this zone, but see if we can
3273			 * grow this zone if it contains deferred pages.
3274			 */
3275			if (deferred_pages_enabled()) {
3276				if (_deferred_grow_zone(zone, order))
3277					goto try_this_zone;
3278			}
3279#endif
3280			/* Checked here to keep the fast path fast */
3281			BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3282			if (alloc_flags & ALLOC_NO_WATERMARKS)
3283				goto try_this_zone;
3284
3285			if (!node_reclaim_enabled() ||
3286			    !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3287				continue;
3288
3289			ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3290			switch (ret) {
3291			case NODE_RECLAIM_NOSCAN:
3292				/* did not scan */
3293				continue;
3294			case NODE_RECLAIM_FULL:
3295				/* scanned but unreclaimable */
3296				continue;
3297			default:
3298				/* did we reclaim enough */
3299				if (zone_watermark_ok(zone, order, mark,
3300					ac->highest_zoneidx, alloc_flags))
3301					goto try_this_zone;
3302
3303				continue;
3304			}
3305		}
3306
3307try_this_zone:
3308		page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3309				gfp_mask, alloc_flags, ac->migratetype);
3310		if (page) {
3311			prep_new_page(page, order, gfp_mask, alloc_flags);
 
 
 
 
3312
3313			/*
3314			 * If this is a high-order atomic allocation then check
3315			 * if the pageblock should be reserved for the future
3316			 */
3317			if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3318				reserve_highatomic_pageblock(page, zone);
 
3319
3320			return page;
3321		} else {
3322			if (has_unaccepted_memory()) {
3323				if (try_to_accept_memory(zone, order))
3324					goto try_this_zone;
3325			}
 
3326
3327#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3328			/* Try again if zone has deferred pages */
3329			if (deferred_pages_enabled()) {
3330				if (_deferred_grow_zone(zone, order))
3331					goto try_this_zone;
3332			}
3333#endif
3334		}
3335	}
3336
3337	/*
3338	 * It's possible on a UMA machine to get through all zones that are
3339	 * fragmented. If avoiding fragmentation, reset and try again.
3340	 */
3341	if (no_fallback) {
3342		alloc_flags &= ~ALLOC_NOFRAGMENT;
3343		goto retry;
3344	}
3345
3346	return NULL;
3347}
 
3348
3349static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3350{
 
3351	unsigned int filter = SHOW_MEM_FILTER_NODES;
3352
 
 
 
3353	/*
3354	 * This documents exceptions given to allocations in certain
3355	 * contexts that are allowed to allocate outside current's set
3356	 * of allowed nodes.
3357	 */
3358	if (!(gfp_mask & __GFP_NOMEMALLOC))
3359		if (tsk_is_oom_victim(current) ||
3360		    (current->flags & (PF_MEMALLOC | PF_EXITING)))
3361			filter &= ~SHOW_MEM_FILTER_NODES;
3362	if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3363		filter &= ~SHOW_MEM_FILTER_NODES;
3364
3365	__show_mem(filter, nodemask, gfp_zone(gfp_mask));
3366}
3367
3368void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3369{
3370	struct va_format vaf;
3371	va_list args;
3372	static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3373
3374	if ((gfp_mask & __GFP_NOWARN) ||
3375	     !__ratelimit(&nopage_rs) ||
3376	     ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3377		return;
3378
3379	va_start(args, fmt);
3380	vaf.fmt = fmt;
3381	vaf.va = &args;
3382	pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3383			current->comm, &vaf, gfp_mask, &gfp_mask,
3384			nodemask_pr_args(nodemask));
3385	va_end(args);
3386
3387	cpuset_print_current_mems_allowed();
3388	pr_cont("\n");
3389	dump_stack();
3390	warn_alloc_show_mem(gfp_mask, nodemask);
 
3391}
3392
3393static inline struct page *
3394__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3395			      unsigned int alloc_flags,
3396			      const struct alloc_context *ac)
3397{
3398	struct page *page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3399
3400	page = get_page_from_freelist(gfp_mask, order,
3401			alloc_flags|ALLOC_CPUSET, ac);
3402	/*
3403	 * fallback to ignore cpuset restriction if our nodes
3404	 * are depleted
3405	 */
3406	if (!page)
3407		page = get_page_from_freelist(gfp_mask, order,
3408				alloc_flags, ac);
3409
3410	return page;
3411}
3412
3413static inline struct page *
3414__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3415	const struct alloc_context *ac, unsigned long *did_some_progress)
 
 
3416{
3417	struct oom_control oc = {
3418		.zonelist = ac->zonelist,
3419		.nodemask = ac->nodemask,
3420		.memcg = NULL,
3421		.gfp_mask = gfp_mask,
3422		.order = order,
3423	};
3424	struct page *page;
3425
3426	*did_some_progress = 0;
3427
3428	/*
3429	 * Acquire the oom lock.  If that fails, somebody else is
3430	 * making progress for us.
3431	 */
3432	if (!mutex_trylock(&oom_lock)) {
3433		*did_some_progress = 1;
3434		schedule_timeout_uninterruptible(1);
3435		return NULL;
3436	}
3437
3438	/*
3439	 * Go through the zonelist yet one more time, keep very high watermark
3440	 * here, this is only to catch a parallel oom killing, we must fail if
3441	 * we're still under heavy pressure. But make sure that this reclaim
3442	 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3443	 * allocation which will never fail due to oom_lock already held.
3444	 */
3445	page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3446				      ~__GFP_DIRECT_RECLAIM, order,
3447				      ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3448	if (page)
3449		goto out;
3450
3451	/* Coredumps can quickly deplete all memory reserves */
3452	if (current->flags & PF_DUMPCORE)
3453		goto out;
3454	/* The OOM killer will not help higher order allocs */
3455	if (order > PAGE_ALLOC_COSTLY_ORDER)
3456		goto out;
3457	/*
3458	 * We have already exhausted all our reclaim opportunities without any
3459	 * success so it is time to admit defeat. We will skip the OOM killer
3460	 * because it is very likely that the caller has a more reasonable
3461	 * fallback than shooting a random task.
3462	 *
3463	 * The OOM killer may not free memory on a specific node.
3464	 */
3465	if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3466		goto out;
3467	/* The OOM killer does not needlessly kill tasks for lowmem */
3468	if (ac->highest_zoneidx < ZONE_NORMAL)
3469		goto out;
3470	if (pm_suspended_storage())
3471		goto out;
3472	/*
3473	 * XXX: GFP_NOFS allocations should rather fail than rely on
3474	 * other request to make a forward progress.
3475	 * We are in an unfortunate situation where out_of_memory cannot
3476	 * do much for this context but let's try it to at least get
3477	 * access to memory reserved if the current task is killed (see
3478	 * out_of_memory). Once filesystems are ready to handle allocation
3479	 * failures more gracefully we should just bail out here.
3480	 */
3481
3482	/* Exhausted what can be done so it's blame time */
3483	if (out_of_memory(&oc) ||
3484	    WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3485		*did_some_progress = 1;
3486
3487		/*
3488		 * Help non-failing allocations by giving them access to memory
3489		 * reserves
 
 
 
3490		 */
3491		if (gfp_mask & __GFP_NOFAIL)
3492			page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3493					ALLOC_NO_WATERMARKS, ac);
3494	}
 
 
 
3495out:
3496	mutex_unlock(&oom_lock);
3497	return page;
3498}
3499
3500/*
3501 * Maximum number of compaction retries with a progress before OOM
3502 * killer is consider as the only way to move forward.
3503 */
3504#define MAX_COMPACT_RETRIES 16
3505
3506#ifdef CONFIG_COMPACTION
3507/* Try memory compaction for high-order allocations before reclaim */
3508static struct page *
3509__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3510		unsigned int alloc_flags, const struct alloc_context *ac,
3511		enum compact_priority prio, enum compact_result *compact_result)
 
 
3512{
3513	struct page *page = NULL;
3514	unsigned long pflags;
3515	unsigned int noreclaim_flag;
3516
3517	if (!order)
3518		return NULL;
3519
3520	psi_memstall_enter(&pflags);
3521	delayacct_compact_start();
3522	noreclaim_flag = memalloc_noreclaim_save();
3523
3524	*compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3525								prio, &page);
3526
3527	memalloc_noreclaim_restore(noreclaim_flag);
3528	psi_memstall_leave(&pflags);
3529	delayacct_compact_end();
3530
3531	if (*compact_result == COMPACT_SKIPPED)
3532		return NULL;
3533	/*
3534	 * At least in one zone compaction wasn't deferred or skipped, so let's
3535	 * count a compaction stall
3536	 */
3537	count_vm_event(COMPACTSTALL);
3538
3539	/* Prep a captured page if available */
3540	if (page)
3541		prep_new_page(page, order, gfp_mask, alloc_flags);
3542
3543	/* Try get a page from the freelist if available */
3544	if (!page)
3545		page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3546
3547	if (page) {
3548		struct zone *zone = page_zone(page);
3549
3550		zone->compact_blockskip_flush = false;
3551		compaction_defer_reset(zone, order, true);
3552		count_vm_event(COMPACTSUCCESS);
3553		return page;
3554	}
3555
3556	/*
3557	 * It's bad if compaction run occurs and fails. The most likely reason
3558	 * is that pages exist, but not enough to satisfy watermarks.
3559	 */
3560	count_vm_event(COMPACTFAIL);
3561
3562	cond_resched();
3563
3564	return NULL;
3565}
3566
3567static inline bool
3568should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3569		     enum compact_result compact_result,
3570		     enum compact_priority *compact_priority,
3571		     int *compaction_retries)
3572{
3573	int max_retries = MAX_COMPACT_RETRIES;
3574	int min_priority;
3575	bool ret = false;
3576	int retries = *compaction_retries;
3577	enum compact_priority priority = *compact_priority;
3578
3579	if (!order)
3580		return false;
3581
3582	if (fatal_signal_pending(current))
3583		return false;
3584
3585	/*
3586	 * Compaction was skipped due to a lack of free order-0
3587	 * migration targets. Continue if reclaim can help.
3588	 */
3589	if (compact_result == COMPACT_SKIPPED) {
3590		ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3591		goto out;
3592	}
3593
3594	/*
3595	 * Compaction managed to coalesce some page blocks, but the
3596	 * allocation failed presumably due to a race. Retry some.
3597	 */
3598	if (compact_result == COMPACT_SUCCESS) {
3599		/*
3600		 * !costly requests are much more important than
3601		 * __GFP_RETRY_MAYFAIL costly ones because they are de
3602		 * facto nofail and invoke OOM killer to move on while
3603		 * costly can fail and users are ready to cope with
3604		 * that. 1/4 retries is rather arbitrary but we would
3605		 * need much more detailed feedback from compaction to
3606		 * make a better decision.
3607		 */
3608		if (order > PAGE_ALLOC_COSTLY_ORDER)
3609			max_retries /= 4;
3610
3611		if (++(*compaction_retries) <= max_retries) {
3612			ret = true;
3613			goto out;
3614		}
3615	}
3616
3617	/*
3618	 * Compaction failed. Retry with increasing priority.
3619	 */
3620	min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3621			MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3622
3623	if (*compact_priority > min_priority) {
3624		(*compact_priority)--;
3625		*compaction_retries = 0;
3626		ret = true;
3627	}
3628out:
3629	trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3630	return ret;
3631}
3632#else
3633static inline struct page *
3634__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3635		unsigned int alloc_flags, const struct alloc_context *ac,
3636		enum compact_priority prio, enum compact_result *compact_result)
 
 
3637{
3638	*compact_result = COMPACT_SKIPPED;
3639	return NULL;
3640}
3641
3642static inline bool
3643should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3644		     enum compact_result compact_result,
3645		     enum compact_priority *compact_priority,
3646		     int *compaction_retries)
3647{
3648	struct zone *zone;
3649	struct zoneref *z;
3650
3651	if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3652		return false;
3653
3654	/*
3655	 * There are setups with compaction disabled which would prefer to loop
3656	 * inside the allocator rather than hit the oom killer prematurely.
3657	 * Let's give them a good hope and keep retrying while the order-0
3658	 * watermarks are OK.
3659	 */
3660	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3661				ac->highest_zoneidx, ac->nodemask) {
3662		if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3663					ac->highest_zoneidx, alloc_flags))
3664			return true;
3665	}
3666	return false;
3667}
3668#endif /* CONFIG_COMPACTION */
3669
3670#ifdef CONFIG_LOCKDEP
3671static struct lockdep_map __fs_reclaim_map =
3672	STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3673
3674static bool __need_reclaim(gfp_t gfp_mask)
 
3675{
3676	/* no reclaim without waiting on it */
3677	if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3678		return false;
3679
3680	/* this guy won't enter reclaim */
3681	if (current->flags & PF_MEMALLOC)
3682		return false;
3683
3684	if (gfp_mask & __GFP_NOLOCKDEP)
3685		return false;
3686
3687	return true;
3688}
3689
3690void __fs_reclaim_acquire(unsigned long ip)
3691{
3692	lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3693}
3694
3695void __fs_reclaim_release(unsigned long ip)
3696{
3697	lock_release(&__fs_reclaim_map, ip);
3698}
3699
3700void fs_reclaim_acquire(gfp_t gfp_mask)
3701{
3702	gfp_mask = current_gfp_context(gfp_mask);
3703
3704	if (__need_reclaim(gfp_mask)) {
3705		if (gfp_mask & __GFP_FS)
3706			__fs_reclaim_acquire(_RET_IP_);
3707
3708#ifdef CONFIG_MMU_NOTIFIER
3709		lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3710		lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3711#endif
3712
3713	}
3714}
3715EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3716
3717void fs_reclaim_release(gfp_t gfp_mask)
3718{
3719	gfp_mask = current_gfp_context(gfp_mask);
3720
3721	if (__need_reclaim(gfp_mask)) {
3722		if (gfp_mask & __GFP_FS)
3723			__fs_reclaim_release(_RET_IP_);
3724	}
3725}
3726EXPORT_SYMBOL_GPL(fs_reclaim_release);
3727#endif
3728
3729/*
3730 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3731 * have been rebuilt so allocation retries. Reader side does not lock and
3732 * retries the allocation if zonelist changes. Writer side is protected by the
3733 * embedded spin_lock.
3734 */
3735static DEFINE_SEQLOCK(zonelist_update_seq);
3736
3737static unsigned int zonelist_iter_begin(void)
3738{
3739	if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3740		return read_seqbegin(&zonelist_update_seq);
3741
3742	return 0;
3743}
3744
3745static unsigned int check_retry_zonelist(unsigned int seq)
3746{
3747	if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3748		return read_seqretry(&zonelist_update_seq, seq);
3749
3750	return seq;
3751}
3752
3753/* Perform direct synchronous page reclaim */
3754static unsigned long
3755__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3756					const struct alloc_context *ac)
3757{
3758	unsigned int noreclaim_flag;
3759	unsigned long progress;
3760
3761	cond_resched();
3762
3763	/* We now go into synchronous reclaim */
3764	cpuset_memory_pressure_bump();
3765	fs_reclaim_acquire(gfp_mask);
3766	noreclaim_flag = memalloc_noreclaim_save();
3767
3768	progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3769								ac->nodemask);
3770
3771	memalloc_noreclaim_restore(noreclaim_flag);
3772	fs_reclaim_release(gfp_mask);
 
 
3773
3774	cond_resched();
3775
3776	return progress;
3777}
3778
3779/* The really slow allocator path where we enter direct reclaim */
3780static inline struct page *
3781__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3782		unsigned int alloc_flags, const struct alloc_context *ac,
3783		unsigned long *did_some_progress)
3784{
3785	struct page *page = NULL;
3786	unsigned long pflags;
3787	bool drained = false;
3788
3789	psi_memstall_enter(&pflags);
3790	*did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3791	if (unlikely(!(*did_some_progress)))
3792		goto out;
3793
3794retry:
3795	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
 
 
 
3796
3797	/*
3798	 * If an allocation failed after direct reclaim, it could be because
3799	 * pages are pinned on the per-cpu lists or in high alloc reserves.
3800	 * Shrink them and try again
3801	 */
3802	if (!page && !drained) {
3803		unreserve_highatomic_pageblock(ac, false);
3804		drain_all_pages(NULL);
3805		drained = true;
3806		goto retry;
3807	}
3808out:
3809	psi_memstall_leave(&pflags);
3810
3811	return page;
3812}
3813
3814static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3815			     const struct alloc_context *ac)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3816{
3817	struct zoneref *z;
3818	struct zone *zone;
3819	pg_data_t *last_pgdat = NULL;
3820	enum zone_type highest_zoneidx = ac->highest_zoneidx;
3821
3822	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3823					ac->nodemask) {
3824		if (!managed_zone(zone))
3825			continue;
3826		if (last_pgdat != zone->zone_pgdat) {
3827			wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3828			last_pgdat = zone->zone_pgdat;
3829		}
3830	}
3831}
3832
3833static inline unsigned int
3834gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3835{
3836	unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
 
3837
3838	/*
3839	 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3840	 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3841	 * to save two branches.
3842	 */
3843	BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3844	BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3845
3846	/*
3847	 * The caller may dip into page reserves a bit more if the caller
3848	 * cannot run direct reclaim, or if the caller has realtime scheduling
3849	 * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
3850	 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3851	 */
3852	alloc_flags |= (__force int)
3853		(gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3854
3855	if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3856		/*
3857		 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3858		 * if it can't schedule.
3859		 */
3860		if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3861			alloc_flags |= ALLOC_NON_BLOCK;
3862
3863			if (order > 0)
3864				alloc_flags |= ALLOC_HIGHATOMIC;
3865		}
3866
3867		/*
3868		 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3869		 * GFP_ATOMIC) rather than fail, see the comment for
3870		 * cpuset_node_allowed().
3871		 */
3872		if (alloc_flags & ALLOC_MIN_RESERVE)
3873			alloc_flags &= ~ALLOC_CPUSET;
3874	} else if (unlikely(rt_task(current)) && in_task())
3875		alloc_flags |= ALLOC_MIN_RESERVE;
3876
3877	alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
 
 
 
 
 
3878
3879	return alloc_flags;
3880}
3881
3882static bool oom_reserves_allowed(struct task_struct *tsk)
3883{
3884	if (!tsk_is_oom_victim(tsk))
3885		return false;
3886
3887	/*
3888	 * !MMU doesn't have oom reaper so give access to memory reserves
3889	 * only to the thread with TIF_MEMDIE set
3890	 */
3891	if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3892		return false;
3893
3894	return true;
3895}
3896
3897/*
3898 * Distinguish requests which really need access to full memory
3899 * reserves from oom victims which can live with a portion of it
3900 */
3901static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3902{
3903	if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3904		return 0;
3905	if (gfp_mask & __GFP_MEMALLOC)
3906		return ALLOC_NO_WATERMARKS;
3907	if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3908		return ALLOC_NO_WATERMARKS;
3909	if (!in_interrupt()) {
3910		if (current->flags & PF_MEMALLOC)
3911			return ALLOC_NO_WATERMARKS;
3912		else if (oom_reserves_allowed(current))
3913			return ALLOC_OOM;
3914	}
3915
3916	return 0;
3917}
3918
3919bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3920{
3921	return !!__gfp_pfmemalloc_flags(gfp_mask);
3922}
3923
3924/*
3925 * Checks whether it makes sense to retry the reclaim to make a forward progress
3926 * for the given allocation request.
3927 *
3928 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3929 * without success, or when we couldn't even meet the watermark if we
3930 * reclaimed all remaining pages on the LRU lists.
3931 *
3932 * Returns true if a retry is viable or false to enter the oom path.
3933 */
3934static inline bool
3935should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3936		     struct alloc_context *ac, int alloc_flags,
3937		     bool did_some_progress, int *no_progress_loops)
3938{
3939	struct zone *zone;
3940	struct zoneref *z;
3941	bool ret = false;
3942
3943	/*
3944	 * Costly allocations might have made a progress but this doesn't mean
3945	 * their order will become available due to high fragmentation so
3946	 * always increment the no progress counter for them
3947	 */
3948	if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3949		*no_progress_loops = 0;
3950	else
3951		(*no_progress_loops)++;
3952
3953	if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3954		goto out;
3955
3956
3957	/*
3958	 * Keep reclaiming pages while there is a chance this will lead
3959	 * somewhere.  If none of the target zones can satisfy our allocation
3960	 * request even if all reclaimable pages are considered then we are
3961	 * screwed and have to go OOM.
3962	 */
3963	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3964				ac->highest_zoneidx, ac->nodemask) {
3965		unsigned long available;
3966		unsigned long reclaimable;
3967		unsigned long min_wmark = min_wmark_pages(zone);
3968		bool wmark;
3969
3970		available = reclaimable = zone_reclaimable_pages(zone);
3971		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3972
3973		/*
3974		 * Would the allocation succeed if we reclaimed all
3975		 * reclaimable pages?
3976		 */
3977		wmark = __zone_watermark_ok(zone, order, min_wmark,
3978				ac->highest_zoneidx, alloc_flags, available);
3979		trace_reclaim_retry_zone(z, order, reclaimable,
3980				available, min_wmark, *no_progress_loops, wmark);
3981		if (wmark) {
3982			ret = true;
3983			break;
3984		}
3985	}
3986
3987	/*
3988	 * Memory allocation/reclaim might be called from a WQ context and the
3989	 * current implementation of the WQ concurrency control doesn't
3990	 * recognize that a particular WQ is congested if the worker thread is
3991	 * looping without ever sleeping. Therefore we have to do a short sleep
3992	 * here rather than calling cond_resched().
3993	 */
3994	if (current->flags & PF_WQ_WORKER)
3995		schedule_timeout_uninterruptible(1);
3996	else
3997		cond_resched();
3998out:
3999	/* Before OOM, exhaust highatomic_reserve */
4000	if (!ret)
4001		return unreserve_highatomic_pageblock(ac, true);
4002
4003	return ret;
4004}
4005
4006static inline bool
4007check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4008{
4009	/*
4010	 * It's possible that cpuset's mems_allowed and the nodemask from
4011	 * mempolicy don't intersect. This should be normally dealt with by
4012	 * policy_nodemask(), but it's possible to race with cpuset update in
4013	 * such a way the check therein was true, and then it became false
4014	 * before we got our cpuset_mems_cookie here.
4015	 * This assumes that for all allocations, ac->nodemask can come only
4016	 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4017	 * when it does not intersect with the cpuset restrictions) or the
4018	 * caller can deal with a violated nodemask.
4019	 */
4020	if (cpusets_enabled() && ac->nodemask &&
4021			!cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4022		ac->nodemask = NULL;
4023		return true;
4024	}
4025
4026	/*
4027	 * When updating a task's mems_allowed or mempolicy nodemask, it is
4028	 * possible to race with parallel threads in such a way that our
4029	 * allocation can fail while the mask is being updated. If we are about
4030	 * to fail, check if the cpuset changed during allocation and if so,
4031	 * retry.
4032	 */
4033	if (read_mems_allowed_retry(cpuset_mems_cookie))
4034		return true;
4035
4036	return false;
4037}
4038
4039static inline struct page *
4040__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4041						struct alloc_context *ac)
 
 
4042{
4043	bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4044	bool can_compact = gfp_compaction_allowed(gfp_mask);
4045	const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4046	struct page *page = NULL;
4047	unsigned int alloc_flags;
 
4048	unsigned long did_some_progress;
4049	enum compact_priority compact_priority;
4050	enum compact_result compact_result;
4051	int compaction_retries;
4052	int no_progress_loops;
4053	unsigned int cpuset_mems_cookie;
4054	unsigned int zonelist_iter_cookie;
4055	int reserve_flags;
4056
4057restart:
4058	compaction_retries = 0;
4059	no_progress_loops = 0;
4060	compact_priority = DEF_COMPACT_PRIORITY;
4061	cpuset_mems_cookie = read_mems_allowed_begin();
4062	zonelist_iter_cookie = zonelist_iter_begin();
4063
4064	/*
4065	 * The fast path uses conservative alloc_flags to succeed only until
4066	 * kswapd needs to be woken up, and to avoid the cost of setting up
4067	 * alloc_flags precisely. So we do that now.
 
4068	 */
4069	alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
 
 
 
4070
4071	/*
4072	 * We need to recalculate the starting point for the zonelist iterator
4073	 * because we might have used different nodemask in the fast path, or
4074	 * there was a cpuset modification and we are retrying - otherwise we
4075	 * could end up iterating over non-eligible zones endlessly.
 
 
4076	 */
4077	ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4078					ac->highest_zoneidx, ac->nodemask);
4079	if (!ac->preferred_zoneref->zone)
4080		goto nopage;
4081
 
 
 
 
 
4082	/*
4083	 * Check for insane configurations where the cpuset doesn't contain
4084	 * any suitable zone to satisfy the request - e.g. non-movable
4085	 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4086	 */
4087	if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4088		struct zoneref *z = first_zones_zonelist(ac->zonelist,
4089					ac->highest_zoneidx,
4090					&cpuset_current_mems_allowed);
4091		if (!z->zone)
4092			goto nopage;
4093	}
4094
4095	if (alloc_flags & ALLOC_KSWAPD)
4096		wake_all_kswapds(order, gfp_mask, ac);
4097
4098	/*
4099	 * The adjusted alloc_flags might result in immediate success, so try
4100	 * that first
4101	 */
4102	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
 
 
 
 
 
 
 
 
4103	if (page)
4104		goto got_pg;
4105
4106	/*
4107	 * For costly allocations, try direct compaction first, as it's likely
4108	 * that we have enough base pages and don't need to reclaim. For non-
4109	 * movable high-order allocations, do that as well, as compaction will
4110	 * try prevent permanent fragmentation by migrating from blocks of the
4111	 * same migratetype.
4112	 * Don't try this for allocations that are allowed to ignore
4113	 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4114	 */
4115	if (can_direct_reclaim && can_compact &&
4116			(costly_order ||
4117			   (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4118			&& !gfp_pfmemalloc_allowed(gfp_mask)) {
4119		page = __alloc_pages_direct_compact(gfp_mask, order,
4120						alloc_flags, ac,
4121						INIT_COMPACT_PRIORITY,
4122						&compact_result);
4123		if (page)
4124			goto got_pg;
4125
4126		/*
4127		 * Checks for costly allocations with __GFP_NORETRY, which
4128		 * includes some THP page fault allocations
4129		 */
4130		if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4131			/*
4132			 * If allocating entire pageblock(s) and compaction
4133			 * failed because all zones are below low watermarks
4134			 * or is prohibited because it recently failed at this
4135			 * order, fail immediately unless the allocator has
4136			 * requested compaction and reclaim retry.
4137			 *
4138			 * Reclaim is
4139			 *  - potentially very expensive because zones are far
4140			 *    below their low watermarks or this is part of very
4141			 *    bursty high order allocations,
4142			 *  - not guaranteed to help because isolate_freepages()
4143			 *    may not iterate over freed pages as part of its
4144			 *    linear scan, and
4145			 *  - unlikely to make entire pageblocks free on its
4146			 *    own.
4147			 */
4148			if (compact_result == COMPACT_SKIPPED ||
4149			    compact_result == COMPACT_DEFERRED)
4150				goto nopage;
4151
4152			/*
4153			 * Looks like reclaim/compaction is worth trying, but
4154			 * sync compaction could be very expensive, so keep
4155			 * using async compaction.
4156			 */
4157			compact_priority = INIT_COMPACT_PRIORITY;
4158		}
4159	}
4160
4161retry:
4162	/* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4163	if (alloc_flags & ALLOC_KSWAPD)
4164		wake_all_kswapds(order, gfp_mask, ac);
4165
4166	reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4167	if (reserve_flags)
4168		alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4169					  (alloc_flags & ALLOC_KSWAPD);
4170
4171	/*
4172	 * Reset the nodemask and zonelist iterators if memory policies can be
4173	 * ignored. These allocations are high priority and system rather than
4174	 * user oriented.
4175	 */
4176	if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4177		ac->nodemask = NULL;
4178		ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4179					ac->highest_zoneidx, ac->nodemask);
4180	}
4181
4182	/* Attempt with potentially adjusted zonelist and alloc_flags */
4183	page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4184	if (page)
4185		goto got_pg;
4186
4187	/* Caller is not willing to reclaim, we can't balance anything */
4188	if (!can_direct_reclaim)
4189		goto nopage;
4190
4191	/* Avoid recursion of direct reclaim */
4192	if (current->flags & PF_MEMALLOC)
4193		goto nopage;
4194
4195	/* Try direct reclaim and then allocating */
4196	page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4197							&did_some_progress);
4198	if (page)
4199		goto got_pg;
4200
4201	/* Try direct compaction and then allocating */
4202	page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4203					compact_priority, &compact_result);
4204	if (page)
4205		goto got_pg;
4206
4207	/* Do not loop if specifically requested */
4208	if (gfp_mask & __GFP_NORETRY)
4209		goto nopage;
4210
4211	/*
4212	 * Do not retry costly high order allocations unless they are
4213	 * __GFP_RETRY_MAYFAIL and we can compact
4214	 */
4215	if (costly_order && (!can_compact ||
4216			     !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4217		goto nopage;
 
 
 
 
 
 
4218
4219	if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4220				 did_some_progress > 0, &no_progress_loops))
4221		goto retry;
4222
4223	/*
4224	 * It doesn't make any sense to retry for the compaction if the order-0
4225	 * reclaim is not able to make any progress because the current
4226	 * implementation of the compaction depends on the sufficient amount
4227	 * of free memory (see __compaction_suitable)
4228	 */
4229	if (did_some_progress > 0 && can_compact &&
4230			should_compact_retry(ac, order, alloc_flags,
4231				compact_result, &compact_priority,
4232				&compaction_retries))
4233		goto retry;
4234
4235
4236	/*
4237	 * Deal with possible cpuset update races or zonelist updates to avoid
4238	 * a unnecessary OOM kill.
4239	 */
4240	if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4241	    check_retry_zonelist(zonelist_iter_cookie))
4242		goto restart;
4243
4244	/* Reclaim has failed us, start killing things */
4245	page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4246	if (page)
4247		goto got_pg;
4248
4249	/* Avoid allocations with no watermarks from looping endlessly */
4250	if (tsk_is_oom_victim(current) &&
4251	    (alloc_flags & ALLOC_OOM ||
4252	     (gfp_mask & __GFP_NOMEMALLOC)))
4253		goto nopage;
4254
4255	/* Retry as long as the OOM killer is making progress */
4256	if (did_some_progress) {
4257		no_progress_loops = 0;
4258		goto retry;
4259	}
4260
4261nopage:
4262	/*
4263	 * Deal with possible cpuset update races or zonelist updates to avoid
4264	 * a unnecessary OOM kill.
4265	 */
4266	if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4267	    check_retry_zonelist(zonelist_iter_cookie))
4268		goto restart;
 
 
 
 
 
 
 
4269
4270	/*
4271	 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4272	 * we always retry
4273	 */
4274	if (gfp_mask & __GFP_NOFAIL) {
4275		/*
4276		 * All existing users of the __GFP_NOFAIL are blockable, so warn
4277		 * of any new users that actually require GFP_NOWAIT
4278		 */
4279		if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4280			goto fail;
 
 
 
 
 
 
4281
4282		/*
4283		 * PF_MEMALLOC request from this context is rather bizarre
4284		 * because we cannot reclaim anything and only can loop waiting
4285		 * for somebody to do a work for us
4286		 */
4287		WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4288
 
 
 
 
 
 
 
4289		/*
4290		 * non failing costly orders are a hard requirement which we
4291		 * are not prepared for much so let's warn about these users
4292		 * so that we can identify them and convert them to something
4293		 * else.
4294		 */
4295		WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4296
4297		/*
4298		 * Help non-failing allocations by giving some access to memory
4299		 * reserves normally used for high priority non-blocking
4300		 * allocations but do not use ALLOC_NO_WATERMARKS because this
4301		 * could deplete whole memory reserves which would just make
4302		 * the situation worse.
4303		 */
4304		page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4305		if (page)
4306			goto got_pg;
 
4307
4308		cond_resched();
4309		goto retry;
4310	}
4311fail:
4312	warn_alloc(gfp_mask, ac->nodemask,
4313			"page allocation failure: order:%u", order);
4314got_pg:
 
 
4315	return page;
4316}
4317
4318static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4319		int preferred_nid, nodemask_t *nodemask,
4320		struct alloc_context *ac, gfp_t *alloc_gfp,
4321		unsigned int *alloc_flags)
4322{
4323	ac->highest_zoneidx = gfp_zone(gfp_mask);
4324	ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4325	ac->nodemask = nodemask;
4326	ac->migratetype = gfp_migratetype(gfp_mask);
4327
4328	if (cpusets_enabled()) {
4329		*alloc_gfp |= __GFP_HARDWALL;
4330		/*
4331		 * When we are in the interrupt context, it is irrelevant
4332		 * to the current task context. It means that any node ok.
4333		 */
4334		if (in_task() && !ac->nodemask)
4335			ac->nodemask = &cpuset_current_mems_allowed;
4336		else
4337			*alloc_flags |= ALLOC_CPUSET;
4338	}
4339
4340	might_alloc(gfp_mask);
4341
4342	if (should_fail_alloc_page(gfp_mask, order))
4343		return false;
4344
4345	*alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4346
4347	/* Dirty zone balancing only done in the fast path */
4348	ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4349
4350	/*
4351	 * The preferred zone is used for statistics but crucially it is
4352	 * also used as the starting point for the zonelist iterator. It
4353	 * may get reset for allocations that ignore memory policies.
4354	 */
4355	ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4356					ac->highest_zoneidx, ac->nodemask);
4357
4358	return true;
4359}
4360
4361/*
4362 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4363 * @gfp: GFP flags for the allocation
4364 * @preferred_nid: The preferred NUMA node ID to allocate from
4365 * @nodemask: Set of nodes to allocate from, may be NULL
4366 * @nr_pages: The number of pages desired on the list or array
4367 * @page_list: Optional list to store the allocated pages
4368 * @page_array: Optional array to store the pages
4369 *
4370 * This is a batched version of the page allocator that attempts to
4371 * allocate nr_pages quickly. Pages are added to page_list if page_list
4372 * is not NULL, otherwise it is assumed that the page_array is valid.
4373 *
4374 * For lists, nr_pages is the number of pages that should be allocated.
4375 *
4376 * For arrays, only NULL elements are populated with pages and nr_pages
4377 * is the maximum number of pages that will be stored in the array.
4378 *
4379 * Returns the number of pages on the list or array.
4380 */
4381unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4382			nodemask_t *nodemask, int nr_pages,
4383			struct list_head *page_list,
4384			struct page **page_array)
4385{
 
 
4386	struct page *page;
4387	unsigned long __maybe_unused UP_flags;
4388	struct zone *zone;
4389	struct zoneref *z;
4390	struct per_cpu_pages *pcp;
4391	struct list_head *pcp_list;
4392	struct alloc_context ac;
4393	gfp_t alloc_gfp;
4394	unsigned int alloc_flags = ALLOC_WMARK_LOW;
4395	int nr_populated = 0, nr_account = 0;
4396
4397	/*
4398	 * Skip populated array elements to determine if any pages need
4399	 * to be allocated before disabling IRQs.
4400	 */
4401	while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4402		nr_populated++;
4403
4404	/* No pages requested? */
4405	if (unlikely(nr_pages <= 0))
4406		goto out;
4407
4408	/* Already populated array? */
4409	if (unlikely(page_array && nr_pages - nr_populated == 0))
4410		goto out;
4411
4412	/* Bulk allocator does not support memcg accounting. */
4413	if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4414		goto failed;
4415
4416	/* Use the single page allocator for one page. */
4417	if (nr_pages - nr_populated == 1)
4418		goto failed;
4419
4420#ifdef CONFIG_PAGE_OWNER
4421	/*
4422	 * PAGE_OWNER may recurse into the allocator to allocate space to
4423	 * save the stack with pagesets.lock held. Releasing/reacquiring
4424	 * removes much of the performance benefit of bulk allocation so
4425	 * force the caller to allocate one page at a time as it'll have
4426	 * similar performance to added complexity to the bulk allocator.
4427	 */
4428	if (static_branch_unlikely(&page_owner_inited))
4429		goto failed;
4430#endif
4431
4432	/* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4433	gfp &= gfp_allowed_mask;
4434	alloc_gfp = gfp;
4435	if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4436		goto out;
4437	gfp = alloc_gfp;
4438
4439	/* Find an allowed local zone that meets the low watermark. */
4440	for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4441		unsigned long mark;
4442
4443		if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4444		    !__cpuset_zone_allowed(zone, gfp)) {
4445			continue;
4446		}
4447
4448		if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4449		    zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4450			goto failed;
4451		}
4452
4453		mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4454		if (zone_watermark_fast(zone, 0,  mark,
4455				zonelist_zone_idx(ac.preferred_zoneref),
4456				alloc_flags, gfp)) {
4457			break;
4458		}
4459	}
4460
4461	/*
4462	 * If there are no allowed local zones that meets the watermarks then
4463	 * try to allocate a single page and reclaim if necessary.
4464	 */
4465	if (unlikely(!zone))
4466		goto failed;
4467
4468	/* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4469	pcp_trylock_prepare(UP_flags);
4470	pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4471	if (!pcp)
4472		goto failed_irq;
4473
4474	/* Attempt the batch allocation */
4475	pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4476	while (nr_populated < nr_pages) {
4477
4478		/* Skip existing pages */
4479		if (page_array && page_array[nr_populated]) {
4480			nr_populated++;
4481			continue;
4482		}
4483
4484		page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4485								pcp, pcp_list);
4486		if (unlikely(!page)) {
4487			/* Try and allocate at least one page */
4488			if (!nr_account) {
4489				pcp_spin_unlock(pcp);
4490				goto failed_irq;
4491			}
4492			break;
4493		}
4494		nr_account++;
4495
4496		prep_new_page(page, 0, gfp, 0);
4497		if (page_list)
4498			list_add(&page->lru, page_list);
4499		else
4500			page_array[nr_populated] = page;
4501		nr_populated++;
4502	}
4503
4504	pcp_spin_unlock(pcp);
4505	pcp_trylock_finish(UP_flags);
4506
4507	__count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4508	zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4509
4510out:
4511	return nr_populated;
4512
4513failed_irq:
4514	pcp_trylock_finish(UP_flags);
4515
4516failed:
4517	page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4518	if (page) {
4519		if (page_list)
4520			list_add(&page->lru, page_list);
4521		else
4522			page_array[nr_populated] = page;
4523		nr_populated++;
4524	}
4525
4526	goto out;
4527}
4528EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4529
4530/*
4531 * This is the 'heart' of the zoned buddy allocator.
4532 */
4533struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4534							nodemask_t *nodemask)
4535{
4536	struct page *page;
4537	unsigned int alloc_flags = ALLOC_WMARK_LOW;
4538	gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4539	struct alloc_context ac = { };
4540
4541	/*
4542	 * There are several places where we assume that the order value is sane
4543	 * so bail out early if the request is out of bound.
 
4544	 */
4545	if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4546		return NULL;
4547
4548	gfp &= gfp_allowed_mask;
4549	/*
4550	 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4551	 * resp. GFP_NOIO which has to be inherited for all allocation requests
4552	 * from a particular context which has been marked by
4553	 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4554	 * movable zones are not used during allocation.
4555	 */
4556	gfp = current_gfp_context(gfp);
4557	alloc_gfp = gfp;
4558	if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4559			&alloc_gfp, &alloc_flags))
4560		return NULL;
4561
4562	/*
4563	 * Forbid the first pass from falling back to types that fragment
4564	 * memory until all local zones are considered.
4565	 */
4566	alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4567
4568	/* First allocation attempt */
4569	page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4570	if (likely(page))
4571		goto out;
4572
4573	alloc_gfp = gfp;
4574	ac.spread_dirty_pages = false;
4575
4576	/*
4577	 * Restore the original nodemask if it was potentially replaced with
4578	 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4579	 */
4580	ac.nodemask = nodemask;
4581
4582	page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4583
4584out:
4585	if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4586	    unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4587		__free_pages(page, order);
4588		page = NULL;
4589	}
4590
4591	trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4592	kmsan_alloc_page(page, order, alloc_gfp);
4593
 
4594	return page;
4595}
4596EXPORT_SYMBOL(__alloc_pages);
4597
4598struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4599		nodemask_t *nodemask)
4600{
4601	struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4602					preferred_nid, nodemask);
4603	return page_rmappable_folio(page);
4604}
4605EXPORT_SYMBOL(__folio_alloc);
4606
4607/*
4608 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4609 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4610 * you need to access high mem.
4611 */
4612unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4613{
4614	struct page *page;
4615
4616	page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
 
 
 
 
 
 
4617	if (!page)
4618		return 0;
4619	return (unsigned long) page_address(page);
4620}
4621EXPORT_SYMBOL(__get_free_pages);
4622
4623unsigned long get_zeroed_page(gfp_t gfp_mask)
4624{
4625	return __get_free_page(gfp_mask | __GFP_ZERO);
4626}
4627EXPORT_SYMBOL(get_zeroed_page);
4628
4629/**
4630 * __free_pages - Free pages allocated with alloc_pages().
4631 * @page: The page pointer returned from alloc_pages().
4632 * @order: The order of the allocation.
4633 *
4634 * This function can free multi-page allocations that are not compound
4635 * pages.  It does not check that the @order passed in matches that of
4636 * the allocation, so it is easy to leak memory.  Freeing more memory
4637 * than was allocated will probably emit a warning.
4638 *
4639 * If the last reference to this page is speculative, it will be released
4640 * by put_page() which only frees the first page of a non-compound
4641 * allocation.  To prevent the remaining pages from being leaked, we free
4642 * the subsequent pages here.  If you want to use the page's reference
4643 * count to decide when to free the allocation, you should allocate a
4644 * compound page, and use put_page() instead of __free_pages().
4645 *
4646 * Context: May be called in interrupt context or while holding a normal
4647 * spinlock, but not in NMI context or while holding a raw spinlock.
4648 */
4649void __free_pages(struct page *page, unsigned int order)
4650{
4651	/* get PageHead before we drop reference */
4652	int head = PageHead(page);
 
 
 
 
 
4653
4654	if (put_page_testzero(page))
4655		free_the_page(page, order);
4656	else if (!head)
4657		while (order-- > 0)
4658			free_the_page(page + (1 << order), order);
4659}
4660EXPORT_SYMBOL(__free_pages);
4661
4662void free_pages(unsigned long addr, unsigned int order)
4663{
4664	if (addr != 0) {
4665		VM_BUG_ON(!virt_addr_valid((void *)addr));
4666		__free_pages(virt_to_page((void *)addr), order);
4667	}
4668}
4669
4670EXPORT_SYMBOL(free_pages);
4671
4672/*
4673 * Page Fragment:
4674 *  An arbitrary-length arbitrary-offset area of memory which resides
4675 *  within a 0 or higher order page.  Multiple fragments within that page
4676 *  are individually refcounted, in the page's reference counter.
4677 *
4678 * The page_frag functions below provide a simple allocation framework for
4679 * page fragments.  This is used by the network stack and network device
4680 * drivers to provide a backing region of memory for use as either an
4681 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4682 */
4683static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4684					     gfp_t gfp_mask)
4685{
4686	struct page *page = NULL;
4687	gfp_t gfp = gfp_mask;
 
4688
4689#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4690	gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4691		    __GFP_NOMEMALLOC;
4692	page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4693				PAGE_FRAG_CACHE_MAX_ORDER);
4694	nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4695#endif
4696	if (unlikely(!page))
4697		page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4698
4699	nc->va = page ? page_address(page) : NULL;
4700
4701	return page;
4702}
4703
4704void __page_frag_cache_drain(struct page *page, unsigned int count)
4705{
4706	VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4707
4708	if (page_ref_sub_and_test(page, count))
4709		free_the_page(page, compound_order(page));
4710}
4711EXPORT_SYMBOL(__page_frag_cache_drain);
4712
4713void *page_frag_alloc_align(struct page_frag_cache *nc,
4714		      unsigned int fragsz, gfp_t gfp_mask,
4715		      unsigned int align_mask)
4716{
4717	unsigned int size = PAGE_SIZE;
4718	struct page *page;
4719	int offset;
4720
4721	if (unlikely(!nc->va)) {
4722refill:
4723		page = __page_frag_cache_refill(nc, gfp_mask);
4724		if (!page)
4725			return NULL;
4726
4727#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4728		/* if size can vary use size else just use PAGE_SIZE */
4729		size = nc->size;
4730#endif
4731		/* Even if we own the page, we do not use atomic_set().
4732		 * This would break get_page_unless_zero() users.
4733		 */
4734		page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4735
4736		/* reset page count bias and offset to start of new frag */
4737		nc->pfmemalloc = page_is_pfmemalloc(page);
4738		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4739		nc->offset = size;
4740	}
4741
4742	offset = nc->offset - fragsz;
4743	if (unlikely(offset < 0)) {
4744		page = virt_to_page(nc->va);
4745
4746		if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4747			goto refill;
4748
4749		if (unlikely(nc->pfmemalloc)) {
4750			free_the_page(page, compound_order(page));
4751			goto refill;
4752		}
4753
4754#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4755		/* if size can vary use size else just use PAGE_SIZE */
4756		size = nc->size;
4757#endif
4758		/* OK, page count is 0, we can safely set it */
4759		set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4760
4761		/* reset page count bias and offset to start of new frag */
4762		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4763		offset = size - fragsz;
4764		if (unlikely(offset < 0)) {
4765			/*
4766			 * The caller is trying to allocate a fragment
4767			 * with fragsz > PAGE_SIZE but the cache isn't big
4768			 * enough to satisfy the request, this may
4769			 * happen in low memory conditions.
4770			 * We don't release the cache page because
4771			 * it could make memory pressure worse
4772			 * so we simply return NULL here.
4773			 */
4774			return NULL;
4775		}
4776	}
4777
4778	nc->pagecnt_bias--;
4779	offset &= align_mask;
4780	nc->offset = offset;
4781
4782	return nc->va + offset;
4783}
4784EXPORT_SYMBOL(page_frag_alloc_align);
4785
4786/*
4787 * Frees a page fragment allocated out of either a compound or order 0 page.
4788 */
4789void page_frag_free(void *addr)
4790{
4791	struct page *page = virt_to_head_page(addr);
4792
4793	if (unlikely(put_page_testzero(page)))
4794		free_the_page(page, compound_order(page));
4795}
4796EXPORT_SYMBOL(page_frag_free);
4797
4798static void *make_alloc_exact(unsigned long addr, unsigned int order,
4799		size_t size)
4800{
4801	if (addr) {
4802		unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4803		struct page *page = virt_to_page((void *)addr);
4804		struct page *last = page + nr;
4805
4806		split_page_owner(page, 1 << order);
4807		split_page_memcg(page, 1 << order);
4808		while (page < --last)
4809			set_page_refcounted(last);
4810
4811		last = page + (1UL << order);
4812		for (page += nr; page < last; page++)
4813			__free_pages_ok(page, 0, FPI_TO_TAIL);
4814	}
4815	return (void *)addr;
4816}
4817
4818/**
4819 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4820 * @size: the number of bytes to allocate
4821 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4822 *
4823 * This function is similar to alloc_pages(), except that it allocates the
4824 * minimum number of pages to satisfy the request.  alloc_pages() can only
4825 * allocate memory in power-of-two pages.
4826 *
4827 * This function is also limited by MAX_PAGE_ORDER.
4828 *
4829 * Memory allocated by this function must be released by free_pages_exact().
4830 *
4831 * Return: pointer to the allocated area or %NULL in case of error.
4832 */
4833void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4834{
4835	unsigned int order = get_order(size);
4836	unsigned long addr;
4837
4838	if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4839		gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4840
4841	addr = __get_free_pages(gfp_mask, order);
4842	return make_alloc_exact(addr, order, size);
4843}
4844EXPORT_SYMBOL(alloc_pages_exact);
4845
4846/**
4847 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4848 *			   pages on a node.
4849 * @nid: the preferred node ID where memory should be allocated
4850 * @size: the number of bytes to allocate
4851 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4852 *
4853 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4854 * back.
4855 *
4856 * Return: pointer to the allocated area or %NULL in case of error.
4857 */
4858void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4859{
4860	unsigned int order = get_order(size);
4861	struct page *p;
4862
4863	if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4864		gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4865
4866	p = alloc_pages_node(nid, gfp_mask, order);
4867	if (!p)
4868		return NULL;
4869	return make_alloc_exact((unsigned long)page_address(p), order, size);
4870}
 
4871
4872/**
4873 * free_pages_exact - release memory allocated via alloc_pages_exact()
4874 * @virt: the value returned by alloc_pages_exact.
4875 * @size: size of allocation, same value as passed to alloc_pages_exact().
4876 *
4877 * Release the memory allocated by a previous call to alloc_pages_exact.
4878 */
4879void free_pages_exact(void *virt, size_t size)
4880{
4881	unsigned long addr = (unsigned long)virt;
4882	unsigned long end = addr + PAGE_ALIGN(size);
4883
4884	while (addr < end) {
4885		free_page(addr);
4886		addr += PAGE_SIZE;
4887	}
4888}
4889EXPORT_SYMBOL(free_pages_exact);
4890
4891/**
4892 * nr_free_zone_pages - count number of pages beyond high watermark
4893 * @offset: The zone index of the highest zone
4894 *
4895 * nr_free_zone_pages() counts the number of pages which are beyond the
4896 * high watermark within all zones at or below a given zone index.  For each
4897 * zone, the number of pages is calculated as:
4898 *
4899 *     nr_free_zone_pages = managed_pages - high_pages
4900 *
4901 * Return: number of pages beyond high watermark.
4902 */
4903static unsigned long nr_free_zone_pages(int offset)
4904{
4905	struct zoneref *z;
4906	struct zone *zone;
4907
4908	/* Just pick one node, since fallback list is circular */
4909	unsigned long sum = 0;
4910
4911	struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4912
4913	for_each_zone_zonelist(zone, z, zonelist, offset) {
4914		unsigned long size = zone_managed_pages(zone);
4915		unsigned long high = high_wmark_pages(zone);
4916		if (size > high)
4917			sum += size - high;
4918	}
4919
4920	return sum;
4921}
4922
4923/**
4924 * nr_free_buffer_pages - count number of pages beyond high watermark
4925 *
4926 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4927 * watermark within ZONE_DMA and ZONE_NORMAL.
4928 *
4929 * Return: number of pages beyond high watermark within ZONE_DMA and
4930 * ZONE_NORMAL.
4931 */
4932unsigned long nr_free_buffer_pages(void)
4933{
4934	return nr_free_zone_pages(gfp_zone(GFP_USER));
4935}
4936EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4937
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4938static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4939{
4940	zoneref->zone = zone;
4941	zoneref->zone_idx = zone_idx(zone);
4942}
4943
4944/*
4945 * Builds allocation fallback zone lists.
4946 *
4947 * Add all populated zones of a node to the zonelist.
4948 */
4949static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
 
4950{
4951	struct zone *zone;
4952	enum zone_type zone_type = MAX_NR_ZONES;
4953	int nr_zones = 0;
 
4954
4955	do {
4956		zone_type--;
4957		zone = pgdat->node_zones + zone_type;
4958		if (populated_zone(zone)) {
4959			zoneref_set_zone(zone, &zonerefs[nr_zones++]);
 
4960			check_highest_zone(zone_type);
4961		}
 
4962	} while (zone_type);
4963
4964	return nr_zones;
4965}
4966
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4967#ifdef CONFIG_NUMA
 
 
 
 
 
 
 
 
 
 
 
 
 
4968
4969static int __parse_numa_zonelist_order(char *s)
4970{
4971	/*
4972	 * We used to support different zonelists modes but they turned
4973	 * out to be just not useful. Let's keep the warning in place
4974	 * if somebody still use the cmd line parameter so that we do
4975	 * not fail it silently
4976	 */
4977	if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4978		pr_warn("Ignoring unsupported numa_zonelist_order value:  %s\n", s);
 
 
4979		return -EINVAL;
4980	}
4981	return 0;
4982}
4983
4984static char numa_zonelist_order[] = "Node";
4985#define NUMA_ZONELIST_ORDER_LEN	16
 
 
 
 
 
 
 
 
 
 
 
 
 
4986/*
4987 * sysctl handler for numa_zonelist_order
4988 */
4989static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4990		void *buffer, size_t *length, loff_t *ppos)
 
4991{
 
 
 
 
 
4992	if (write)
4993		return __parse_numa_zonelist_order(buffer);
4994	return proc_dostring(table, write, buffer, length, ppos);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4995}
4996
 
 
4997static int node_load[MAX_NUMNODES];
4998
4999/**
5000 * find_next_best_node - find the next node that should appear in a given node's fallback list
5001 * @node: node whose fallback list we're appending
5002 * @used_node_mask: nodemask_t of already used nodes
5003 *
5004 * We use a number of factors to determine which is the next node that should
5005 * appear on a given node's fallback list.  The node should not have appeared
5006 * already in @node's fallback list, and it should be the next closest node
5007 * according to the distance array (which contains arbitrary distance values
5008 * from each node to each node in the system), and should also prefer nodes
5009 * with no CPUs, since presumably they'll have very little allocation pressure
5010 * on them otherwise.
5011 *
5012 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5013 */
5014int find_next_best_node(int node, nodemask_t *used_node_mask)
5015{
5016	int n, val;
5017	int min_val = INT_MAX;
5018	int best_node = NUMA_NO_NODE;
 
5019
5020	/*
5021	 * Use the local node if we haven't already, but for memoryless local
5022	 * node, we should skip it and fall back to other nodes.
5023	 */
5024	if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5025		node_set(node, *used_node_mask);
5026		return node;
5027	}
5028
5029	for_each_node_state(n, N_MEMORY) {
5030
5031		/* Don't want a node to appear more than once */
5032		if (node_isset(n, *used_node_mask))
5033			continue;
5034
5035		/* Use the distance array to find the distance */
5036		val = node_distance(node, n);
5037
5038		/* Penalize nodes under us ("prefer the next node") */
5039		val += (n < node);
5040
5041		/* Give preference to headless and unused nodes */
5042		if (!cpumask_empty(cpumask_of_node(n)))
 
5043			val += PENALTY_FOR_NODE_WITH_CPUS;
5044
5045		/* Slight preference for less loaded node */
5046		val *= MAX_NUMNODES;
5047		val += node_load[n];
5048
5049		if (val < min_val) {
5050			min_val = val;
5051			best_node = n;
5052		}
5053	}
5054
5055	if (best_node >= 0)
5056		node_set(best_node, *used_node_mask);
5057
5058	return best_node;
5059}
5060
5061
5062/*
5063 * Build zonelists ordered by node and zones within node.
5064 * This results in maximum locality--normal zone overflows into local
5065 * DMA zone, if any--but risks exhausting DMA zone.
5066 */
5067static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5068		unsigned nr_nodes)
5069{
5070	struct zoneref *zonerefs;
5071	int i;
5072
5073	zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5074
5075	for (i = 0; i < nr_nodes; i++) {
5076		int nr_zones;
5077
5078		pg_data_t *node = NODE_DATA(node_order[i]);
5079
5080		nr_zones = build_zonerefs_node(node, zonerefs);
5081		zonerefs += nr_zones;
5082	}
5083	zonerefs->zone = NULL;
5084	zonerefs->zone_idx = 0;
 
 
5085}
5086
5087/*
5088 * Build gfp_thisnode zonelists
5089 */
5090static void build_thisnode_zonelists(pg_data_t *pgdat)
5091{
5092	struct zoneref *zonerefs;
5093	int nr_zones;
5094
5095	zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5096	nr_zones = build_zonerefs_node(pgdat, zonerefs);
5097	zonerefs += nr_zones;
5098	zonerefs->zone = NULL;
5099	zonerefs->zone_idx = 0;
5100}
5101
5102/*
5103 * Build zonelists ordered by zone and nodes within zones.
5104 * This results in conserving DMA zone[s] until all Normal memory is
5105 * exhausted, but results in overflowing to remote node while memory
5106 * may still exist in local DMA zone.
5107 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5108
5109static void build_zonelists(pg_data_t *pgdat)
5110{
5111	static int node_order[MAX_NUMNODES];
5112	int node, nr_nodes = 0;
5113	nodemask_t used_mask = NODE_MASK_NONE;
5114	int local_node, prev_node;
 
 
 
 
 
 
 
 
 
5115
5116	/* NUMA-aware ordering of nodes */
5117	local_node = pgdat->node_id;
 
5118	prev_node = local_node;
 
5119
5120	memset(node_order, 0, sizeof(node_order));
 
 
5121	while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
 
 
 
 
 
 
 
 
 
5122		/*
5123		 * We don't want to pressure a particular node.
5124		 * So adding penalty to the first node in same
5125		 * distance group to make it round-robin.
5126		 */
5127		if (node_distance(local_node, node) !=
5128		    node_distance(local_node, prev_node))
5129			node_load[node] += 1;
5130
5131		node_order[nr_nodes++] = node;
5132		prev_node = node;
 
 
 
 
 
 
 
 
 
 
5133	}
5134
5135	build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5136	build_thisnode_zonelists(pgdat);
5137	pr_info("Fallback order for Node %d: ", local_node);
5138	for (node = 0; node < nr_nodes; node++)
5139		pr_cont("%d ", node_order[node]);
5140	pr_cont("\n");
 
 
 
 
 
 
 
 
 
 
5141}
5142
5143#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5144/*
5145 * Return node id of node used for "local" allocations.
5146 * I.e., first node id of first zone in arg node's generic zonelist.
5147 * Used for initializing percpu 'numa_mem', which is used primarily
5148 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5149 */
5150int local_memory_node(int node)
5151{
5152	struct zoneref *z;
5153
5154	z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5155				   gfp_zone(GFP_KERNEL),
5156				   NULL);
5157	return zone_to_nid(z->zone);
 
5158}
5159#endif
5160
5161static void setup_min_unmapped_ratio(void);
5162static void setup_min_slab_ratio(void);
5163#else	/* CONFIG_NUMA */
5164
 
 
 
 
 
5165static void build_zonelists(pg_data_t *pgdat)
5166{
5167	int node, local_node;
5168	struct zoneref *zonerefs;
5169	int nr_zones;
5170
5171	local_node = pgdat->node_id;
5172
5173	zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5174	nr_zones = build_zonerefs_node(pgdat, zonerefs);
5175	zonerefs += nr_zones;
5176
5177	/*
5178	 * Now we build the zonelist so that it contains the zones
5179	 * of all the other nodes.
5180	 * We don't want to pressure a particular node, so when
5181	 * building the zones for node N, we make sure that the
5182	 * zones coming right after the local ones are those from
5183	 * node N+1 (modulo N)
5184	 */
5185	for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5186		if (!node_online(node))
5187			continue;
5188		nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5189		zonerefs += nr_zones;
5190	}
5191	for (node = 0; node < local_node; node++) {
5192		if (!node_online(node))
5193			continue;
5194		nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5195		zonerefs += nr_zones;
5196	}
5197
5198	zonerefs->zone = NULL;
5199	zonerefs->zone_idx = 0;
 
 
 
 
 
 
5200}
5201
5202#endif	/* CONFIG_NUMA */
5203
5204/*
5205 * Boot pageset table. One per cpu which is going to be used for all
5206 * zones and all nodes. The parameters will be set in such a way
5207 * that an item put on a list will immediately be handed over to
5208 * the buddy list. This is safe since pageset manipulation is done
5209 * with interrupts disabled.
5210 *
5211 * The boot_pagesets must be kept even after bootup is complete for
5212 * unused processors and/or zones. They do play a role for bootstrapping
5213 * hotplugged processors.
5214 *
5215 * zoneinfo_show() and maybe other functions do
5216 * not check if the processor is online before following the pageset pointer.
5217 * Other parts of the kernel may not check if the zone is available.
5218 */
5219static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5220/* These effectively disable the pcplists in the boot pageset completely */
5221#define BOOT_PAGESET_HIGH	0
5222#define BOOT_PAGESET_BATCH	1
5223static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5224static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5225
5226static void __build_all_zonelists(void *data)
 
 
 
 
 
 
 
5227{
5228	int nid;
5229	int __maybe_unused cpu;
5230	pg_data_t *self = data;
5231	unsigned long flags;
5232
5233	/*
5234	 * The zonelist_update_seq must be acquired with irqsave because the
5235	 * reader can be invoked from IRQ with GFP_ATOMIC.
5236	 */
5237	write_seqlock_irqsave(&zonelist_update_seq, flags);
5238	/*
5239	 * Also disable synchronous printk() to prevent any printk() from
5240	 * trying to hold port->lock, for
5241	 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5242	 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5243	 */
5244	printk_deferred_enter();
5245
5246#ifdef CONFIG_NUMA
5247	memset(node_load, 0, sizeof(node_load));
5248#endif
 
 
 
 
 
 
5249
5250	/*
5251	 * This node is hotadded and no memory is yet present.   So just
5252	 * building zonelists is fine - no need to touch other nodes.
 
 
 
 
 
 
 
 
 
5253	 */
5254	if (self && !node_online(self->node_id)) {
5255		build_zonelists(self);
5256	} else {
5257		/*
5258		 * All possible nodes have pgdat preallocated
5259		 * in free_area_init
5260		 */
5261		for_each_node(nid) {
5262			pg_data_t *pgdat = NODE_DATA(nid);
5263
5264			build_zonelists(pgdat);
5265		}
5266
5267#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5268		/*
5269		 * We now know the "local memory node" for each node--
5270		 * i.e., the node of the first zone in the generic zonelist.
5271		 * Set up numa_mem percpu variable for on-line cpus.  During
5272		 * boot, only the boot cpu should be on-line;  we'll init the
5273		 * secondary cpus' numa_mem as they come on-line.  During
5274		 * node/memory hotplug, we'll fixup all on-line cpus.
5275		 */
5276		for_each_online_cpu(cpu)
5277			set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5278#endif
5279	}
5280
5281	printk_deferred_exit();
5282	write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5283}
5284
5285static noinline void __init
5286build_all_zonelists_init(void)
5287{
5288	int cpu;
5289
5290	__build_all_zonelists(NULL);
5291
5292	/*
5293	 * Initialize the boot_pagesets that are going to be used
5294	 * for bootstrapping processors. The real pagesets for
5295	 * each zone will be allocated later when the per cpu
5296	 * allocator is available.
5297	 *
5298	 * boot_pagesets are used also for bootstrapping offline
5299	 * cpus if the system is already booted because the pagesets
5300	 * are needed to initialize allocators on a specific cpu too.
5301	 * F.e. the percpu allocator needs the page allocator which
5302	 * needs the percpu allocator in order to allocate its pagesets
5303	 * (a chicken-egg dilemma).
5304	 */
5305	for_each_possible_cpu(cpu)
5306		per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5307
5308	mminit_verify_zonelist();
5309	cpuset_init_current_mems_allowed();
5310}
5311
5312/*
 
5313 * unless system_state == SYSTEM_BOOTING.
5314 *
5315 * __ref due to call of __init annotated helper build_all_zonelists_init
5316 * [protected by SYSTEM_BOOTING].
5317 */
5318void __ref build_all_zonelists(pg_data_t *pgdat)
5319{
5320	unsigned long vm_total_pages;
5321
5322	if (system_state == SYSTEM_BOOTING) {
5323		build_all_zonelists_init();
 
 
5324	} else {
5325		__build_all_zonelists(pgdat);
 
 
 
 
 
 
5326		/* cpuset refresh routine should be here */
5327	}
5328	/* Get the number of free pages beyond high watermark in all zones. */
5329	vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5330	/*
5331	 * Disable grouping by mobility if the number of pages in the
5332	 * system is too low to allow the mechanism to work. It would be
5333	 * more accurate, but expensive to check per-zone. This check is
5334	 * made on memory-hotadd so a system can start with mobility
5335	 * disabled and enable it later
5336	 */
5337	if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5338		page_group_by_mobility_disabled = 1;
5339	else
5340		page_group_by_mobility_disabled = 0;
5341
5342	pr_info("Built %u zonelists, mobility grouping %s.  Total pages: %ld\n",
5343		nr_online_nodes,
5344		page_group_by_mobility_disabled ? "off" : "on",
5345		vm_total_pages);
 
 
5346#ifdef CONFIG_NUMA
5347	pr_info("Policy zone: %s\n", zone_names[policy_zone]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5348#endif
 
 
 
 
 
 
 
 
 
 
5349}
5350
 
 
 
 
 
5351static int zone_batchsize(struct zone *zone)
5352{
5353#ifdef CONFIG_MMU
5354	int batch;
5355
5356	/*
5357	 * The number of pages to batch allocate is either ~0.1%
5358	 * of the zone or 1MB, whichever is smaller. The batch
5359	 * size is striking a balance between allocation latency
5360	 * and zone lock contention.
5361	 */
5362	batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
 
 
5363	batch /= 4;		/* We effectively *= 4 below */
5364	if (batch < 1)
5365		batch = 1;
5366
5367	/*
5368	 * Clamp the batch to a 2^n - 1 value. Having a power
5369	 * of 2 value was found to be more likely to have
5370	 * suboptimal cache aliasing properties in some cases.
5371	 *
5372	 * For example if 2 tasks are alternately allocating
5373	 * batches of pages, one task can end up with a lot
5374	 * of pages of one half of the possible page colors
5375	 * and the other with pages of the other colors.
5376	 */
5377	batch = rounddown_pow_of_two(batch + batch/2) - 1;
5378
5379	return batch;
5380
5381#else
5382	/* The deferral and batching of frees should be suppressed under NOMMU
5383	 * conditions.
5384	 *
5385	 * The problem is that NOMMU needs to be able to allocate large chunks
5386	 * of contiguous memory as there's no hardware page translation to
5387	 * assemble apparent contiguous memory from discontiguous pages.
5388	 *
5389	 * Queueing large contiguous runs of pages for batching, however,
5390	 * causes the pages to actually be freed in smaller chunks.  As there
5391	 * can be a significant delay between the individual batches being
5392	 * recycled, this leads to the once large chunks of space being
5393	 * fragmented and becoming unavailable for high-order allocations.
5394	 */
5395	return 0;
5396#endif
5397}
5398
5399static int percpu_pagelist_high_fraction;
5400static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5401			 int high_fraction)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5402{
5403#ifdef CONFIG_MMU
5404	int high;
5405	int nr_split_cpus;
5406	unsigned long total_pages;
 
 
 
 
 
 
 
 
5407
5408	if (!high_fraction) {
5409		/*
5410		 * By default, the high value of the pcp is based on the zone
5411		 * low watermark so that if they are full then background
5412		 * reclaim will not be started prematurely.
5413		 */
5414		total_pages = low_wmark_pages(zone);
 
 
 
 
 
 
 
5415	} else {
5416		/*
5417		 * If percpu_pagelist_high_fraction is configured, the high
5418		 * value is based on a fraction of the managed pages in the
5419		 * zone.
 
 
 
 
 
5420		 */
5421		total_pages = zone_managed_pages(zone) / high_fraction;
5422	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5423
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5424	/*
5425	 * Split the high value across all online CPUs local to the zone. Note
5426	 * that early in boot that CPUs may not be online yet and that during
5427	 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5428	 * onlined. For memory nodes that have no CPUs, split the high value
5429	 * across all online CPUs to mitigate the risk that reclaim is triggered
5430	 * prematurely due to pages stored on pcp lists.
5431	 */
5432	nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5433	if (!nr_split_cpus)
5434		nr_split_cpus = num_online_cpus();
5435	high = total_pages / nr_split_cpus;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5436
5437	/*
5438	 * Ensure high is at least batch*4. The multiple is based on the
5439	 * historical relationship between high and batch.
5440	 */
5441	high = max(high, batch << 2);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5442
5443	return high;
5444#else
 
 
5445	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
5446#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5447}
5448
 
5449/*
5450 * pcp->high and pcp->batch values are related and generally batch is lower
5451 * than high. They are also related to pcp->count such that count is lower
5452 * than high, and as soon as it reaches high, the pcplist is flushed.
5453 *
5454 * However, guaranteeing these relations at all times would require e.g. write
5455 * barriers here but also careful usage of read barriers at the read side, and
5456 * thus be prone to error and bad for performance. Thus the update only prevents
5457 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5458 * should ensure they can cope with those fields changing asynchronously, and
5459 * fully trust only the pcp->count field on the local CPU with interrupts
5460 * disabled.
5461 *
5462 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5463 * outside of boot time (or some other assurance that no concurrent updaters
5464 * exist).
5465 */
5466static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5467			   unsigned long high_max, unsigned long batch)
5468{
5469	WRITE_ONCE(pcp->batch, batch);
5470	WRITE_ONCE(pcp->high_min, high_min);
5471	WRITE_ONCE(pcp->high_max, high_max);
 
 
 
 
5472}
5473
5474static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
 
 
 
 
5475{
5476	int pindex;
 
 
5477
5478	memset(pcp, 0, sizeof(*pcp));
5479	memset(pzstats, 0, sizeof(*pzstats));
 
 
 
 
 
 
 
 
 
5480
5481	spin_lock_init(&pcp->lock);
5482	for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5483		INIT_LIST_HEAD(&pcp->lists[pindex]);
 
 
 
 
 
 
 
5484
5485	/*
5486	 * Set batch and high values safe for a boot pageset. A true percpu
5487	 * pageset's initialization will update them subsequently. Here we don't
5488	 * need to be as careful as pageset_update() as nobody can access the
5489	 * pageset yet.
5490	 */
5491	pcp->high_min = BOOT_PAGESET_HIGH;
5492	pcp->high_max = BOOT_PAGESET_HIGH;
5493	pcp->batch = BOOT_PAGESET_BATCH;
5494	pcp->free_count = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5495}
5496
5497static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5498					      unsigned long high_max, unsigned long batch)
 
 
 
 
 
 
 
 
 
 
 
5499{
5500	struct per_cpu_pages *pcp;
5501	int cpu;
 
5502
5503	for_each_possible_cpu(cpu) {
5504		pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5505		pageset_update(pcp, high_min, high_max, batch);
5506	}
 
 
 
5507}
5508
5509/*
5510 * Calculate and set new high and batch values for all per-cpu pagesets of a
5511 * zone based on the zone's size.
 
5512 */
5513static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5514{
5515	int new_high_min, new_high_max, new_batch;
 
 
 
 
 
 
 
 
 
 
 
 
5516
5517	new_batch = max(1, zone_batchsize(zone));
5518	if (percpu_pagelist_high_fraction) {
5519		new_high_min = zone_highsize(zone, new_batch, cpu_online,
5520					     percpu_pagelist_high_fraction);
5521		/*
5522		 * PCP high is tuned manually, disable auto-tuning via
5523		 * setting high_min and high_max to the manual value.
5524		 */
5525		new_high_max = new_high_min;
5526	} else {
5527		new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5528		new_high_max = zone_highsize(zone, new_batch, cpu_online,
5529					     MIN_PERCPU_PAGELIST_HIGH_FRACTION);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5530	}
 
5531
5532	if (zone->pageset_high_min == new_high_min &&
5533	    zone->pageset_high_max == new_high_max &&
5534	    zone->pageset_batch == new_batch)
5535		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5536
5537	zone->pageset_high_min = new_high_min;
5538	zone->pageset_high_max = new_high_max;
5539	zone->pageset_batch = new_batch;
5540
5541	__zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5542					  new_batch);
5543}
5544
5545void __meminit setup_zone_pageset(struct zone *zone)
 
 
 
 
 
 
5546{
5547	int cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5548
5549	/* Size may be 0 on !SMP && !NUMA */
5550	if (sizeof(struct per_cpu_zonestat) > 0)
5551		zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5552
5553	zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5554	for_each_possible_cpu(cpu) {
5555		struct per_cpu_pages *pcp;
5556		struct per_cpu_zonestat *pzstats;
5557
5558		pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5559		pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5560		per_cpu_pages_init(pcp, pzstats);
 
 
 
5561	}
5562
5563	zone_set_pageset_high_and_batch(zone, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5564}
5565
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5566/*
5567 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5568 * page high values need to be recalculated.
 
 
 
5569 */
5570static void zone_pcp_update(struct zone *zone, int cpu_online)
5571{
5572	mutex_lock(&pcp_batch_high_lock);
5573	zone_set_pageset_high_and_batch(zone, cpu_online);
5574	mutex_unlock(&pcp_batch_high_lock);
 
 
 
 
 
5575}
5576
5577static void zone_pcp_update_cacheinfo(struct zone *zone)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5578{
5579	int cpu;
5580	struct per_cpu_pages *pcp;
5581	struct cpu_cacheinfo *cci;
5582
5583	for_each_online_cpu(cpu) {
5584		pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5585		cci = get_cpu_cacheinfo(cpu);
5586		/*
5587		 * If data cache slice of CPU is large enough, "pcp->batch"
5588		 * pages can be preserved in PCP before draining PCP for
5589		 * consecutive high-order pages freeing without allocation.
5590		 * This can reduce zone lock contention without hurting
5591		 * cache-hot pages sharing.
5592		 */
5593		spin_lock(&pcp->lock);
5594		if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5595			pcp->flags |= PCPF_FREE_HIGH_BATCH;
5596		else
5597			pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5598		spin_unlock(&pcp->lock);
5599	}
5600}
5601
5602void setup_pcp_cacheinfo(void)
 
5603{
5604	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
5605
5606	for_each_populated_zone(zone)
5607		zone_pcp_update_cacheinfo(zone);
 
 
 
 
 
 
 
5608}
 
 
 
5609
5610/*
5611 * Allocate per cpu pagesets and initialize them.
5612 * Before this call only boot pagesets were available.
 
 
5613 */
5614void __init setup_per_cpu_pageset(void)
 
5615{
5616	struct pglist_data *pgdat;
5617	struct zone *zone;
5618	int __maybe_unused cpu;
 
5619
5620	for_each_populated_zone(zone)
5621		setup_zone_pageset(zone);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5622
 
 
5623#ifdef CONFIG_NUMA
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5624	/*
5625	 * Unpopulated zones continue using the boot pagesets.
5626	 * The numa stats for these pagesets need to be reset.
5627	 * Otherwise, they will end up skewing the stats of
5628	 * the nodes these zones are associated with.
5629	 */
5630	for_each_possible_cpu(cpu) {
5631		struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5632		memset(pzstats->vm_numa_event, 0,
5633		       sizeof(pzstats->vm_numa_event));
 
 
5634	}
5635#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5636
5637	for_each_online_pgdat(pgdat)
5638		pgdat->per_cpu_nodestats =
5639			alloc_percpu(struct per_cpu_nodestat);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5640}
5641
5642__meminit void zone_pcp_init(struct zone *zone)
 
5643{
5644	/*
5645	 * per cpu subsystem is not up at this point. The following code
5646	 * relies on the ability of the linker to provide the
5647	 * offset of a (static) per cpu variable into the per cpu area.
5648	 */
5649	zone->per_cpu_pageset = &boot_pageset;
5650	zone->per_cpu_zonestats = &boot_zonestats;
5651	zone->pageset_high_min = BOOT_PAGESET_HIGH;
5652	zone->pageset_high_max = BOOT_PAGESET_HIGH;
5653	zone->pageset_batch = BOOT_PAGESET_BATCH;
5654
5655	if (populated_zone(zone))
5656		pr_debug("  %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5657			 zone->present_pages, zone_batchsize(zone));
5658}
5659
5660void adjust_managed_page_count(struct page *page, long count)
 
5661{
5662	atomic_long_add(count, &page_zone(page)->managed_pages);
5663	totalram_pages_add(count);
5664#ifdef CONFIG_HIGHMEM
5665	if (PageHighMem(page))
5666		totalhigh_pages_add(count);
5667#endif
5668}
5669EXPORT_SYMBOL(adjust_managed_page_count);
5670
5671unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5672{
5673	void *pos;
5674	unsigned long pages = 0;
 
5675
5676	start = (void *)PAGE_ALIGN((unsigned long)start);
5677	end = (void *)((unsigned long)end & PAGE_MASK);
5678	for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5679		struct page *page = virt_to_page(pos);
5680		void *direct_map_addr;
 
 
 
 
 
 
5681
5682		/*
5683		 * 'direct_map_addr' might be different from 'pos'
5684		 * because some architectures' virt_to_page()
5685		 * work with aliases.  Getting the direct map
5686		 * address ensures that we get a _writeable_
5687		 * alias for the memset().
5688		 */
5689		direct_map_addr = page_address(page);
5690		/*
5691		 * Perform a kasan-unchecked memset() since this memory
5692		 * has not been initialized.
5693		 */
5694		direct_map_addr = kasan_reset_tag(direct_map_addr);
5695		if ((unsigned int)poison <= 0xFF)
5696			memset(direct_map_addr, poison, PAGE_SIZE);
 
 
 
 
 
 
 
 
 
5697
5698		free_reserved_page(page);
 
 
 
 
 
 
 
5699	}
5700
5701	if (pages && s)
5702		pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5703
5704	return pages;
 
 
 
 
 
 
 
 
5705}
5706
5707static int page_alloc_cpu_dead(unsigned int cpu)
 
 
 
 
 
5708{
5709	struct zone *zone;
 
 
 
 
 
 
 
 
 
 
 
5710
5711	lru_add_drain_cpu(cpu);
5712	mlock_drain_remote(cpu);
5713	drain_pages(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
5714
5715	/*
5716	 * Spill the event counters of the dead processor
5717	 * into the current processors event counters.
5718	 * This artificially elevates the count of the current
5719	 * processor.
 
 
5720	 */
5721	vm_events_fold_cpu(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5722
5723	/*
5724	 * Zero the differential counters of the dead processor
5725	 * so that the vm statistics are consistent.
5726	 *
5727	 * This is only okay since the processor is dead and cannot
5728	 * race with what we are doing.
5729	 */
5730	cpu_vm_stats_fold(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5731
5732	for_each_populated_zone(zone)
5733		zone_pcp_update(zone, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5734
5735	return 0;
5736}
5737
5738static int page_alloc_cpu_online(unsigned int cpu)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5739{
5740	struct zone *zone;
 
5741
5742	for_each_populated_zone(zone)
5743		zone_pcp_update(zone, 1);
5744	return 0;
 
5745}
5746
5747void __init page_alloc_init_cpuhp(void)
 
5748{
5749	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5750
5751	ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5752					"mm/page_alloc:pcp",
5753					page_alloc_cpu_online,
5754					page_alloc_cpu_dead);
5755	WARN_ON(ret < 0);
5756}
5757
5758/*
5759 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5760 *	or min_free_kbytes changes.
5761 */
5762static void calculate_totalreserve_pages(void)
5763{
5764	struct pglist_data *pgdat;
5765	unsigned long reserve_pages = 0;
5766	enum zone_type i, j;
5767
5768	for_each_online_pgdat(pgdat) {
5769
5770		pgdat->totalreserve_pages = 0;
5771
5772		for (i = 0; i < MAX_NR_ZONES; i++) {
5773			struct zone *zone = pgdat->node_zones + i;
5774			long max = 0;
5775			unsigned long managed_pages = zone_managed_pages(zone);
5776
5777			/* Find valid and maximum lowmem_reserve in the zone */
5778			for (j = i; j < MAX_NR_ZONES; j++) {
5779				if (zone->lowmem_reserve[j] > max)
5780					max = zone->lowmem_reserve[j];
5781			}
5782
5783			/* we treat the high watermark as reserved pages. */
5784			max += high_wmark_pages(zone);
5785
5786			if (max > managed_pages)
5787				max = managed_pages;
5788
5789			pgdat->totalreserve_pages += max;
5790
5791			reserve_pages += max;
5792		}
5793	}
5794	totalreserve_pages = reserve_pages;
5795}
5796
5797/*
5798 * setup_per_zone_lowmem_reserve - called whenever
5799 *	sysctl_lowmem_reserve_ratio changes.  Ensures that each zone
5800 *	has a correct pages reserved value, so an adequate number of
5801 *	pages are left in the zone after a successful __alloc_pages().
5802 */
5803static void setup_per_zone_lowmem_reserve(void)
5804{
5805	struct pglist_data *pgdat;
5806	enum zone_type i, j;
5807
5808	for_each_online_pgdat(pgdat) {
5809		for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5810			struct zone *zone = &pgdat->node_zones[i];
5811			int ratio = sysctl_lowmem_reserve_ratio[i];
5812			bool clear = !ratio || !zone_managed_pages(zone);
5813			unsigned long managed_pages = 0;
5814
5815			for (j = i + 1; j < MAX_NR_ZONES; j++) {
5816				struct zone *upper_zone = &pgdat->node_zones[j];
5817
5818				managed_pages += zone_managed_pages(upper_zone);
5819
5820				if (clear)
5821					zone->lowmem_reserve[j] = 0;
5822				else
5823					zone->lowmem_reserve[j] = managed_pages / ratio;
 
 
 
 
5824			}
5825		}
5826	}
5827
5828	/* update totalreserve_pages */
5829	calculate_totalreserve_pages();
5830}
5831
5832static void __setup_per_zone_wmarks(void)
 
 
 
 
 
 
 
5833{
5834	unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5835	unsigned long lowmem_pages = 0;
5836	struct zone *zone;
5837	unsigned long flags;
5838
5839	/* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5840	for_each_zone(zone) {
5841		if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5842			lowmem_pages += zone_managed_pages(zone);
5843	}
5844
5845	for_each_zone(zone) {
5846		u64 tmp;
5847
5848		spin_lock_irqsave(&zone->lock, flags);
5849		tmp = (u64)pages_min * zone_managed_pages(zone);
5850		do_div(tmp, lowmem_pages);
5851		if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5852			/*
5853			 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5854			 * need highmem and movable zones pages, so cap pages_min
5855			 * to a small  value here.
5856			 *
5857			 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5858			 * deltas control async page reclaim, and so should
5859			 * not be capped for highmem and movable zones.
5860			 */
5861			unsigned long min_pages;
5862
5863			min_pages = zone_managed_pages(zone) / 1024;
5864			min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5865			zone->_watermark[WMARK_MIN] = min_pages;
 
 
 
5866		} else {
5867			/*
5868			 * If it's a lowmem zone, reserve a number of pages
5869			 * proportionate to the zone's size.
5870			 */
5871			zone->_watermark[WMARK_MIN] = tmp;
5872		}
5873
5874		/*
5875		 * Set the kswapd watermarks distance according to the
5876		 * scale factor in proportion to available memory, but
5877		 * ensure a minimum size on small systems.
5878		 */
5879		tmp = max_t(u64, tmp >> 2,
5880			    mult_frac(zone_managed_pages(zone),
5881				      watermark_scale_factor, 10000));
5882
5883		zone->watermark_boost = 0;
5884		zone->_watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
5885		zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5886		zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5887
5888		spin_unlock_irqrestore(&zone->lock, flags);
5889	}
5890
5891	/* update totalreserve_pages */
5892	calculate_totalreserve_pages();
5893}
5894
5895/**
5896 * setup_per_zone_wmarks - called when min_free_kbytes changes
5897 * or when memory is hot-{added|removed}
5898 *
5899 * Ensures that the watermark[min,low,high] values for each zone are set
5900 * correctly with respect to min_free_kbytes.
5901 */
5902void setup_per_zone_wmarks(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5903{
5904	struct zone *zone;
5905	static DEFINE_SPINLOCK(lock);
5906
5907	spin_lock(&lock);
5908	__setup_per_zone_wmarks();
5909	spin_unlock(&lock);
5910
5911	/*
5912	 * The watermark size have changed so update the pcpu batch
5913	 * and high limits or the limits may be inappropriate.
5914	 */
5915	for_each_zone(zone)
5916		zone_pcp_update(zone, 0);
5917}
5918
5919/*
5920 * Initialise min_free_kbytes.
5921 *
5922 * For small machines we want it small (128k min).  For large machines
5923 * we want it large (256MB max).  But it is not linear, because network
5924 * bandwidth does not increase linearly with machine size.  We use
5925 *
5926 *	min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5927 *	min_free_kbytes = sqrt(lowmem_kbytes * 16)
5928 *
5929 * which yields
5930 *
5931 * 16MB:	512k
5932 * 32MB:	724k
5933 * 64MB:	1024k
5934 * 128MB:	1448k
5935 * 256MB:	2048k
5936 * 512MB:	2896k
5937 * 1024MB:	4096k
5938 * 2048MB:	5792k
5939 * 4096MB:	8192k
5940 * 8192MB:	11584k
5941 * 16384MB:	16384k
5942 */
5943void calculate_min_free_kbytes(void)
5944{
5945	unsigned long lowmem_kbytes;
5946	int new_min_free_kbytes;
5947
5948	lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5949	new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5950
5951	if (new_min_free_kbytes > user_min_free_kbytes)
5952		min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5953	else
5954		pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5955				new_min_free_kbytes, user_min_free_kbytes);
5956
5957}
5958
5959int __meminit init_per_zone_wmark_min(void)
5960{
5961	calculate_min_free_kbytes();
 
 
5962	setup_per_zone_wmarks();
5963	refresh_zone_stat_thresholds();
5964	setup_per_zone_lowmem_reserve();
5965
5966#ifdef CONFIG_NUMA
5967	setup_min_unmapped_ratio();
5968	setup_min_slab_ratio();
5969#endif
5970
5971	khugepaged_min_free_kbytes_update();
5972
5973	return 0;
5974}
5975postcore_initcall(init_per_zone_wmark_min)
5976
5977/*
5978 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5979 *	that we can call two helper functions whenever min_free_kbytes
5980 *	changes.
5981 */
5982static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5983		void *buffer, size_t *length, loff_t *ppos)
5984{
5985	int rc;
5986
5987	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5988	if (rc)
5989		return rc;
5990
5991	if (write) {
5992		user_min_free_kbytes = min_free_kbytes;
5993		setup_per_zone_wmarks();
5994	}
5995	return 0;
5996}
5997
5998static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5999		void *buffer, size_t *length, loff_t *ppos)
6000{
6001	int rc;
6002
6003	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6004	if (rc)
6005		return rc;
6006
6007	if (write)
6008		setup_per_zone_wmarks();
6009
6010	return 0;
6011}
6012
6013#ifdef CONFIG_NUMA
6014static void setup_min_unmapped_ratio(void)
 
6015{
6016	pg_data_t *pgdat;
6017	struct zone *zone;
6018
6019	for_each_online_pgdat(pgdat)
6020		pgdat->min_unmapped_pages = 0;
6021
6022	for_each_zone(zone)
6023		zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6024						         sysctl_min_unmapped_ratio) / 100;
6025}
6026
6027
6028static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6029		void *buffer, size_t *length, loff_t *ppos)
6030{
6031	int rc;
6032
6033	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6034	if (rc)
6035		return rc;
6036
6037	setup_min_unmapped_ratio();
6038
 
6039	return 0;
6040}
6041
6042static void setup_min_slab_ratio(void)
 
6043{
6044	pg_data_t *pgdat;
6045	struct zone *zone;
6046
6047	for_each_online_pgdat(pgdat)
6048		pgdat->min_slab_pages = 0;
6049
6050	for_each_zone(zone)
6051		zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6052						     sysctl_min_slab_ratio) / 100;
6053}
6054
6055static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6056		void *buffer, size_t *length, loff_t *ppos)
6057{
6058	int rc;
6059
6060	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6061	if (rc)
6062		return rc;
6063
6064	setup_min_slab_ratio();
6065
 
6066	return 0;
6067}
6068#endif
6069
6070/*
6071 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6072 *	proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6073 *	whenever sysctl_lowmem_reserve_ratio changes.
6074 *
6075 * The reserve ratio obviously has absolutely no relation with the
6076 * minimum watermarks. The lowmem reserve ratio can only make sense
6077 * if in function of the boot time zone sizes.
6078 */
6079static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6080		int write, void *buffer, size_t *length, loff_t *ppos)
6081{
6082	int i;
6083
6084	proc_dointvec_minmax(table, write, buffer, length, ppos);
6085
6086	for (i = 0; i < MAX_NR_ZONES; i++) {
6087		if (sysctl_lowmem_reserve_ratio[i] < 1)
6088			sysctl_lowmem_reserve_ratio[i] = 0;
6089	}
6090
6091	setup_per_zone_lowmem_reserve();
6092	return 0;
6093}
6094
6095/*
6096 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6097 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6098 * pagelist can have before it gets flushed back to buddy allocator.
6099 */
6100static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6101		int write, void *buffer, size_t *length, loff_t *ppos)
 
6102{
6103	struct zone *zone;
6104	int old_percpu_pagelist_high_fraction;
6105	int ret;
6106
6107	mutex_lock(&pcp_batch_high_lock);
6108	old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6109
6110	ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6111	if (!write || ret < 0)
6112		goto out;
6113
6114	/* Sanity checking to avoid pcp imbalance */
6115	if (percpu_pagelist_high_fraction &&
6116	    percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6117		percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6118		ret = -EINVAL;
6119		goto out;
6120	}
 
 
6121
6122	/* No change? */
6123	if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6124		goto out;
6125
6126	for_each_populated_zone(zone)
6127		zone_set_pageset_high_and_batch(zone, 0);
6128out:
6129	mutex_unlock(&pcp_batch_high_lock);
6130	return ret;
6131}
6132
6133static struct ctl_table page_alloc_sysctl_table[] = {
6134	{
6135		.procname	= "min_free_kbytes",
6136		.data		= &min_free_kbytes,
6137		.maxlen		= sizeof(min_free_kbytes),
6138		.mode		= 0644,
6139		.proc_handler	= min_free_kbytes_sysctl_handler,
6140		.extra1		= SYSCTL_ZERO,
6141	},
6142	{
6143		.procname	= "watermark_boost_factor",
6144		.data		= &watermark_boost_factor,
6145		.maxlen		= sizeof(watermark_boost_factor),
6146		.mode		= 0644,
6147		.proc_handler	= proc_dointvec_minmax,
6148		.extra1		= SYSCTL_ZERO,
6149	},
6150	{
6151		.procname	= "watermark_scale_factor",
6152		.data		= &watermark_scale_factor,
6153		.maxlen		= sizeof(watermark_scale_factor),
6154		.mode		= 0644,
6155		.proc_handler	= watermark_scale_factor_sysctl_handler,
6156		.extra1		= SYSCTL_ONE,
6157		.extra2		= SYSCTL_THREE_THOUSAND,
6158	},
6159	{
6160		.procname	= "percpu_pagelist_high_fraction",
6161		.data		= &percpu_pagelist_high_fraction,
6162		.maxlen		= sizeof(percpu_pagelist_high_fraction),
6163		.mode		= 0644,
6164		.proc_handler	= percpu_pagelist_high_fraction_sysctl_handler,
6165		.extra1		= SYSCTL_ZERO,
6166	},
6167	{
6168		.procname	= "lowmem_reserve_ratio",
6169		.data		= &sysctl_lowmem_reserve_ratio,
6170		.maxlen		= sizeof(sysctl_lowmem_reserve_ratio),
6171		.mode		= 0644,
6172		.proc_handler	= lowmem_reserve_ratio_sysctl_handler,
6173	},
6174#ifdef CONFIG_NUMA
6175	{
6176		.procname	= "numa_zonelist_order",
6177		.data		= &numa_zonelist_order,
6178		.maxlen		= NUMA_ZONELIST_ORDER_LEN,
6179		.mode		= 0644,
6180		.proc_handler	= numa_zonelist_order_handler,
6181	},
6182	{
6183		.procname	= "min_unmapped_ratio",
6184		.data		= &sysctl_min_unmapped_ratio,
6185		.maxlen		= sizeof(sysctl_min_unmapped_ratio),
6186		.mode		= 0644,
6187		.proc_handler	= sysctl_min_unmapped_ratio_sysctl_handler,
6188		.extra1		= SYSCTL_ZERO,
6189		.extra2		= SYSCTL_ONE_HUNDRED,
6190	},
6191	{
6192		.procname	= "min_slab_ratio",
6193		.data		= &sysctl_min_slab_ratio,
6194		.maxlen		= sizeof(sysctl_min_slab_ratio),
6195		.mode		= 0644,
6196		.proc_handler	= sysctl_min_slab_ratio_sysctl_handler,
6197		.extra1		= SYSCTL_ZERO,
6198		.extra2		= SYSCTL_ONE_HUNDRED,
6199	},
6200#endif
6201	{}
6202};
6203
6204void __init page_alloc_sysctl_init(void)
6205{
6206	register_sysctl_init("vm", page_alloc_sysctl_table);
 
 
 
6207}
 
 
6208
6209#ifdef CONFIG_CONTIG_ALLOC
6210/* Usage: See admin-guide/dynamic-debug-howto.rst */
6211static void alloc_contig_dump_pages(struct list_head *page_list)
6212{
6213	DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6214
6215	if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6216		struct page *page;
 
 
 
 
 
 
 
 
 
 
6217
6218		dump_stack();
6219		list_for_each_entry(page, page_list, lru)
6220			dump_page(page, "migration failure");
 
6221	}
6222}
6223
6224/* [start, end) must belong to a single zone. */
6225int __alloc_contig_migrate_range(struct compact_control *cc,
6226					unsigned long start, unsigned long end)
6227{
6228	/* This function is based on compact_zone() from compaction.c. */
6229	unsigned int nr_reclaimed;
6230	unsigned long pfn = start;
6231	unsigned int tries = 0;
6232	int ret = 0;
6233	struct migration_target_control mtc = {
6234		.nid = zone_to_nid(cc->zone),
6235		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6236	};
6237
6238	lru_cache_disable();
6239
6240	while (pfn < end || !list_empty(&cc->migratepages)) {
6241		if (fatal_signal_pending(current)) {
6242			ret = -EINTR;
6243			break;
 
6244		}
 
6245
6246		if (list_empty(&cc->migratepages)) {
6247			cc->nr_migratepages = 0;
6248			ret = isolate_migratepages_range(cc, pfn, end);
6249			if (ret && ret != -EAGAIN)
6250				break;
6251			pfn = cc->migrate_pfn;
6252			tries = 0;
6253		} else if (++tries == 5) {
6254			ret = -EBUSY;
6255			break;
6256		}
6257
6258		nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6259							&cc->migratepages);
6260		cc->nr_migratepages -= nr_reclaimed;
 
 
 
 
 
 
 
6261
6262		ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6263			NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6264
6265		/*
6266		 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6267		 * to retry again over this error, so do the same here.
6268		 */
6269		if (ret == -ENOMEM)
6270			break;
6271	}
 
 
 
6272
6273	lru_cache_enable();
6274	if (ret < 0) {
6275		if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6276			alloc_contig_dump_pages(&cc->migratepages);
6277		putback_movable_pages(&cc->migratepages);
6278		return ret;
6279	}
6280	return 0;
 
6281}
6282
6283/**
6284 * alloc_contig_range() -- tries to allocate given range of pages
6285 * @start:	start PFN to allocate
6286 * @end:	one-past-the-last PFN to allocate
6287 * @migratetype:	migratetype of the underlying pageblocks (either
6288 *			#MIGRATE_MOVABLE or #MIGRATE_CMA).  All pageblocks
6289 *			in range must have the same migratetype and it must
6290 *			be either of the two.
6291 * @gfp_mask:	GFP mask to use during compaction
6292 *
6293 * The PFN range does not have to be pageblock aligned. The PFN range must
6294 * belong to a single zone.
6295 *
6296 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6297 * pageblocks in the range.  Once isolated, the pageblocks should not
6298 * be modified by others.
6299 *
6300 * Return: zero on success or negative error code.  On success all
6301 * pages which PFN is in [start, end) are allocated for the caller and
6302 * need to be freed with free_contig_range().
6303 */
6304int alloc_contig_range(unsigned long start, unsigned long end,
6305		       unsigned migratetype, gfp_t gfp_mask)
6306{
6307	unsigned long outer_start, outer_end;
6308	int order;
6309	int ret = 0;
 
 
6310
6311	struct compact_control cc = {
6312		.nr_migratepages = 0,
6313		.order = -1,
6314		.zone = page_zone(pfn_to_page(start)),
6315		.mode = MIGRATE_SYNC,
6316		.ignore_skip_hint = true,
6317		.no_set_skip_hint = true,
6318		.gfp_mask = current_gfp_context(gfp_mask),
6319		.alloc_contig = true,
6320	};
6321	INIT_LIST_HEAD(&cc.migratepages);
6322
6323	/*
6324	 * What we do here is we mark all pageblocks in range as
6325	 * MIGRATE_ISOLATE.  Because pageblock and max order pages may
6326	 * have different sizes, and due to the way page allocator
6327	 * work, start_isolate_page_range() has special handlings for this.
6328	 *
6329	 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6330	 * migrate the pages from an unaligned range (ie. pages that
6331	 * we are interested in). This will put all the pages in
6332	 * range back to page allocator as MIGRATE_ISOLATE.
6333	 *
6334	 * When this is done, we take the pages in range from page
6335	 * allocator removing them from the buddy system.  This way
6336	 * page allocator will never consider using them.
6337	 *
6338	 * This lets us mark the pageblocks back as
6339	 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6340	 * aligned range but not in the unaligned, original range are
6341	 * put back to page allocator so that buddy can use them.
6342	 */
6343
6344	ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6345	if (ret)
6346		goto done;
6347
6348	drain_all_pages(cc.zone);
6349
6350	/*
6351	 * In case of -EBUSY, we'd like to know which page causes problem.
6352	 * So, just fall through. test_pages_isolated() has a tracepoint
6353	 * which will report the busy page.
6354	 *
6355	 * It is possible that busy pages could become available before
6356	 * the call to test_pages_isolated, and the range will actually be
6357	 * allocated.  So, if we fall through be sure to clear ret so that
6358	 * -EBUSY is not accidentally used or returned to caller.
6359	 */
6360	ret = __alloc_contig_migrate_range(&cc, start, end);
6361	if (ret && ret != -EBUSY)
6362		goto done;
6363	ret = 0;
6364
6365	/*
6366	 * Pages from [start, end) are within a pageblock_nr_pages
6367	 * aligned blocks that are marked as MIGRATE_ISOLATE.  What's
6368	 * more, all pages in [start, end) are free in page allocator.
6369	 * What we are going to do is to allocate all pages from
6370	 * [start, end) (that is remove them from page allocator).
6371	 *
6372	 * The only problem is that pages at the beginning and at the
6373	 * end of interesting range may be not aligned with pages that
6374	 * page allocator holds, ie. they can be part of higher order
6375	 * pages.  Because of this, we reserve the bigger range and
6376	 * once this is done free the pages we are not interested in.
6377	 *
6378	 * We don't have to hold zone->lock here because the pages are
6379	 * isolated thus they won't get removed from buddy.
6380	 */
6381
6382	order = 0;
6383	outer_start = start;
6384	while (!PageBuddy(pfn_to_page(outer_start))) {
6385		if (++order > MAX_PAGE_ORDER) {
6386			outer_start = start;
6387			break;
6388		}
6389		outer_start &= ~0UL << order;
6390	}
6391
6392	if (outer_start != start) {
6393		order = buddy_order(pfn_to_page(outer_start));
6394
6395		/*
6396		 * outer_start page could be small order buddy page and
6397		 * it doesn't include start page. Adjust outer_start
6398		 * in this case to report failed page properly
6399		 * on tracepoint in test_pages_isolated()
6400		 */
6401		if (outer_start + (1UL << order) <= start)
6402			outer_start = start;
6403	}
6404
6405	/* Make sure the range is really isolated. */
6406	if (test_pages_isolated(outer_start, end, 0)) {
6407		ret = -EBUSY;
6408		goto done;
6409	}
6410
6411	/* Grab isolated pages from freelists. */
6412	outer_end = isolate_freepages_range(&cc, outer_start, end);
6413	if (!outer_end) {
6414		ret = -EBUSY;
6415		goto done;
6416	}
6417
6418	/* Free head and tail (if any) */
6419	if (start != outer_start)
6420		free_contig_range(outer_start, start - outer_start);
6421	if (end != outer_end)
6422		free_contig_range(end, outer_end - end);
6423
6424done:
6425	undo_isolate_page_range(start, end, migratetype);
6426	return ret;
6427}
6428EXPORT_SYMBOL(alloc_contig_range);
6429
6430static int __alloc_contig_pages(unsigned long start_pfn,
6431				unsigned long nr_pages, gfp_t gfp_mask)
 
 
 
 
 
 
 
6432{
6433	unsigned long end_pfn = start_pfn + nr_pages;
 
 
 
6434
6435	return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6436				  gfp_mask);
 
 
 
 
 
 
 
 
 
 
6437}
6438
6439static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6440				   unsigned long nr_pages)
 
 
 
 
 
 
6441{
6442	unsigned long i, end_pfn = start_pfn + nr_pages;
6443	struct page *page;
 
 
 
 
 
6444
6445	for (i = start_pfn; i < end_pfn; i++) {
6446		page = pfn_to_online_page(i);
6447		if (!page)
6448			return false;
6449
6450		if (page_zone(page) != z)
6451			return false;
 
6452
6453		if (PageReserved(page))
6454			return false;
6455
6456		if (PageHuge(page))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6457			return false;
6458	}
6459	return true;
6460}
6461
6462static bool zone_spans_last_pfn(const struct zone *zone,
6463				unsigned long start_pfn, unsigned long nr_pages)
6464{
6465	unsigned long last_pfn = start_pfn + nr_pages - 1;
6466
6467	return zone_spans_pfn(zone, last_pfn);
6468}
6469
6470/**
6471 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6472 * @nr_pages:	Number of contiguous pages to allocate
6473 * @gfp_mask:	GFP mask to limit search and used during compaction
6474 * @nid:	Target node
6475 * @nodemask:	Mask for other possible nodes
6476 *
6477 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6478 * on an applicable zonelist to find a contiguous pfn range which can then be
6479 * tried for allocation with alloc_contig_range(). This routine is intended
6480 * for allocation requests which can not be fulfilled with the buddy allocator.
6481 *
6482 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6483 * power of two, then allocated range is also guaranteed to be aligned to same
6484 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6485 *
6486 * Allocated pages can be freed with free_contig_range() or by manually calling
6487 * __free_page() on each allocated page.
6488 *
6489 * Return: pointer to contiguous pages on success, or NULL if not successful.
6490 */
6491struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6492				int nid, nodemask_t *nodemask)
6493{
6494	unsigned long ret, pfn, flags;
6495	struct zonelist *zonelist;
6496	struct zone *zone;
6497	struct zoneref *z;
 
 
 
6498
6499	zonelist = node_zonelist(nid, gfp_mask);
6500	for_each_zone_zonelist_nodemask(zone, z, zonelist,
6501					gfp_zone(gfp_mask), nodemask) {
6502		spin_lock_irqsave(&zone->lock, flags);
6503
6504		pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6505		while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6506			if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6507				/*
6508				 * We release the zone lock here because
6509				 * alloc_contig_range() will also lock the zone
6510				 * at some point. If there's an allocation
6511				 * spinning on this lock, it may win the race
6512				 * and cause alloc_contig_range() to fail...
6513				 */
6514				spin_unlock_irqrestore(&zone->lock, flags);
6515				ret = __alloc_contig_pages(pfn, nr_pages,
6516							gfp_mask);
6517				if (!ret)
6518					return pfn_to_page(pfn);
6519				spin_lock_irqsave(&zone->lock, flags);
6520			}
6521			pfn += nr_pages;
6522		}
6523		spin_unlock_irqrestore(&zone->lock, flags);
6524	}
6525	return NULL;
6526}
6527#endif /* CONFIG_CONTIG_ALLOC */
6528
6529void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6530{
6531	unsigned long count = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6532
6533	for (; nr_pages--; pfn++) {
6534		struct page *page = pfn_to_page(pfn);
 
 
6535
6536		count += page_count(page) != 1;
6537		__free_page(page);
 
 
6538	}
6539	WARN(count != 0, "%lu pages are still in use!\n", count);
6540}
6541EXPORT_SYMBOL(free_contig_range);
6542
6543/*
6544 * Effectively disable pcplists for the zone by setting the high limit to 0
6545 * and draining all cpus. A concurrent page freeing on another CPU that's about
6546 * to put the page on pcplist will either finish before the drain and the page
6547 * will be drained, or observe the new high limit and skip the pcplist.
6548 *
6549 * Must be paired with a call to zone_pcp_enable().
6550 */
6551void zone_pcp_disable(struct zone *zone)
6552{
6553	mutex_lock(&pcp_batch_high_lock);
6554	__zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6555	__drain_all_pages(zone, true);
6556}
6557
6558void zone_pcp_enable(struct zone *zone)
6559{
6560	__zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6561		zone->pageset_high_max, zone->pageset_batch);
6562	mutex_unlock(&pcp_batch_high_lock);
6563}
6564
6565void zone_pcp_reset(struct zone *zone)
6566{
6567	int cpu;
6568	struct per_cpu_zonestat *pzstats;
6569
6570	if (zone->per_cpu_pageset != &boot_pageset) {
6571		for_each_online_cpu(cpu) {
6572			pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6573			drain_zonestat(zone, pzstats);
6574		}
6575		free_percpu(zone->per_cpu_pageset);
6576		zone->per_cpu_pageset = &boot_pageset;
6577		if (zone->per_cpu_zonestats != &boot_zonestats) {
6578			free_percpu(zone->per_cpu_zonestats);
6579			zone->per_cpu_zonestats = &boot_zonestats;
6580		}
6581	}
6582}
6583
6584#ifdef CONFIG_MEMORY_HOTREMOVE
6585/*
6586 * All pages in the range must be in a single zone, must not contain holes,
6587 * must span full sections, and must be isolated before calling this function.
6588 */
6589void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 
6590{
6591	unsigned long pfn = start_pfn;
6592	struct page *page;
6593	struct zone *zone;
6594	unsigned int order;
 
6595	unsigned long flags;
6596
6597	offline_mem_sections(pfn, end_pfn);
 
 
 
 
6598	zone = page_zone(pfn_to_page(pfn));
6599	spin_lock_irqsave(&zone->lock, flags);
 
6600	while (pfn < end_pfn) {
6601		page = pfn_to_page(pfn);
6602		/*
6603		 * The HWPoisoned page may be not in buddy system, and
6604		 * page_count() is not 0.
6605		 */
6606		if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6607			pfn++;
6608			continue;
6609		}
6610		/*
6611		 * At this point all remaining PageOffline() pages have a
6612		 * reference count of 0 and can simply be skipped.
6613		 */
6614		if (PageOffline(page)) {
6615			BUG_ON(page_count(page));
6616			BUG_ON(PageBuddy(page));
6617			pfn++;
6618			continue;
6619		}
6620
6621		BUG_ON(page_count(page));
6622		BUG_ON(!PageBuddy(page));
6623		order = buddy_order(page);
6624		del_page_from_free_list(page, zone, order);
 
 
 
 
 
 
 
 
 
 
6625		pfn += (1 << order);
6626	}
6627	spin_unlock_irqrestore(&zone->lock, flags);
6628}
6629#endif
6630
6631/*
6632 * This function returns a stable result only if called under zone lock.
6633 */
6634bool is_free_buddy_page(struct page *page)
6635{
6636	unsigned long pfn = page_to_pfn(page);
6637	unsigned int order;
6638
6639	for (order = 0; order < NR_PAGE_ORDERS; order++) {
6640		struct page *page_head = page - (pfn & ((1 << order) - 1));
6641
6642		if (PageBuddy(page_head) &&
6643		    buddy_order_unsafe(page_head) >= order)
6644			break;
6645	}
6646
6647	return order <= MAX_PAGE_ORDER;
6648}
6649EXPORT_SYMBOL(is_free_buddy_page);
6650
6651#ifdef CONFIG_MEMORY_FAILURE
6652/*
6653 * Break down a higher-order page in sub-pages, and keep our target out of
6654 * buddy allocator.
6655 */
6656static void break_down_buddy_pages(struct zone *zone, struct page *page,
6657				   struct page *target, int low, int high,
6658				   int migratetype)
6659{
6660	unsigned long size = 1 << high;
6661	struct page *current_buddy;
6662
6663	while (high > low) {
6664		high--;
6665		size >>= 1;
6666
6667		if (target >= &page[size]) {
6668			current_buddy = page;
6669			page = page + size;
6670		} else {
6671			current_buddy = page + size;
6672		}
6673
6674		if (set_page_guard(zone, current_buddy, high, migratetype))
6675			continue;
6676
6677		add_to_free_list(current_buddy, zone, high, migratetype);
6678		set_buddy_order(current_buddy, high);
6679	}
6680}
6681
6682/*
6683 * Take a page that will be marked as poisoned off the buddy allocator.
6684 */
6685bool take_page_off_buddy(struct page *page)
6686{
6687	struct zone *zone = page_zone(page);
6688	unsigned long pfn = page_to_pfn(page);
6689	unsigned long flags;
6690	unsigned int order;
6691	bool ret = false;
6692
6693	spin_lock_irqsave(&zone->lock, flags);
6694	for (order = 0; order < NR_PAGE_ORDERS; order++) {
6695		struct page *page_head = page - (pfn & ((1 << order) - 1));
6696		int page_order = buddy_order(page_head);
6697
6698		if (PageBuddy(page_head) && page_order >= order) {
6699			unsigned long pfn_head = page_to_pfn(page_head);
6700			int migratetype = get_pfnblock_migratetype(page_head,
6701								   pfn_head);
6702
6703			del_page_from_free_list(page_head, zone, page_order);
6704			break_down_buddy_pages(zone, page_head, page, 0,
6705						page_order, migratetype);
6706			SetPageHWPoisonTakenOff(page);
6707			if (!is_migrate_isolate(migratetype))
6708				__mod_zone_freepage_state(zone, -1, migratetype);
6709			ret = true;
6710			break;
6711		}
6712		if (page_count(page_head) > 0)
6713			break;
6714	}
6715	spin_unlock_irqrestore(&zone->lock, flags);
6716	return ret;
6717}
6718
6719/*
6720 * Cancel takeoff done by take_page_off_buddy().
6721 */
6722bool put_page_back_buddy(struct page *page)
6723{
6724	struct zone *zone = page_zone(page);
6725	unsigned long pfn = page_to_pfn(page);
6726	unsigned long flags;
6727	int migratetype = get_pfnblock_migratetype(page, pfn);
6728	bool ret = false;
6729
6730	spin_lock_irqsave(&zone->lock, flags);
6731	if (put_page_testzero(page)) {
6732		ClearPageHWPoisonTakenOff(page);
6733		__free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6734		if (TestClearPageHWPoison(page)) {
6735			ret = true;
6736		}
6737	}
6738	spin_unlock_irqrestore(&zone->lock, flags);
6739
6740	return ret;
6741}
6742#endif
6743
6744#ifdef CONFIG_ZONE_DMA
6745bool has_managed_dma(void)
6746{
6747	struct pglist_data *pgdat;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6748
6749	for_each_online_pgdat(pgdat) {
6750		struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6751
6752		if (managed_zone(zone))
6753			return true;
6754	}
6755	return false;
6756}
6757#endif /* CONFIG_ZONE_DMA */
6758
6759#ifdef CONFIG_UNACCEPTED_MEMORY
6760
6761/* Counts number of zones with unaccepted pages. */
6762static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6763
6764static bool lazy_accept = true;
6765
6766static int __init accept_memory_parse(char *p)
6767{
6768	if (!strcmp(p, "lazy")) {
6769		lazy_accept = true;
6770		return 0;
6771	} else if (!strcmp(p, "eager")) {
6772		lazy_accept = false;
6773		return 0;
6774	} else {
6775		return -EINVAL;
6776	}
6777}
6778early_param("accept_memory", accept_memory_parse);
6779
6780static bool page_contains_unaccepted(struct page *page, unsigned int order)
6781{
6782	phys_addr_t start = page_to_phys(page);
6783	phys_addr_t end = start + (PAGE_SIZE << order);
6784
6785	return range_contains_unaccepted_memory(start, end);
6786}
6787
6788static void accept_page(struct page *page, unsigned int order)
6789{
6790	phys_addr_t start = page_to_phys(page);
6791
6792	accept_memory(start, start + (PAGE_SIZE << order));
6793}
6794
6795static bool try_to_accept_memory_one(struct zone *zone)
6796{
6797	unsigned long flags;
6798	struct page *page;
6799	bool last;
6800
6801	if (list_empty(&zone->unaccepted_pages))
6802		return false;
6803
6804	spin_lock_irqsave(&zone->lock, flags);
6805	page = list_first_entry_or_null(&zone->unaccepted_pages,
6806					struct page, lru);
6807	if (!page) {
6808		spin_unlock_irqrestore(&zone->lock, flags);
6809		return false;
6810	}
6811
6812	list_del(&page->lru);
6813	last = list_empty(&zone->unaccepted_pages);
6814
6815	__mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6816	__mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6817	spin_unlock_irqrestore(&zone->lock, flags);
6818
6819	accept_page(page, MAX_PAGE_ORDER);
6820
6821	__free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6822
6823	if (last)
6824		static_branch_dec(&zones_with_unaccepted_pages);
6825
6826	return true;
6827}
6828
6829static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6830{
6831	long to_accept;
6832	int ret = false;
6833
6834	/* How much to accept to get to high watermark? */
6835	to_accept = high_wmark_pages(zone) -
6836		    (zone_page_state(zone, NR_FREE_PAGES) -
6837		    __zone_watermark_unusable_free(zone, order, 0));
6838
6839	/* Accept at least one page */
6840	do {
6841		if (!try_to_accept_memory_one(zone))
6842			break;
6843		ret = true;
6844		to_accept -= MAX_ORDER_NR_PAGES;
6845	} while (to_accept > 0);
6846
6847	return ret;
6848}
6849
6850static inline bool has_unaccepted_memory(void)
6851{
6852	return static_branch_unlikely(&zones_with_unaccepted_pages);
6853}
6854
6855static bool __free_unaccepted(struct page *page)
6856{
6857	struct zone *zone = page_zone(page);
6858	unsigned long flags;
6859	bool first = false;
6860
6861	if (!lazy_accept)
6862		return false;
6863
6864	spin_lock_irqsave(&zone->lock, flags);
6865	first = list_empty(&zone->unaccepted_pages);
6866	list_add_tail(&page->lru, &zone->unaccepted_pages);
6867	__mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6868	__mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6869	spin_unlock_irqrestore(&zone->lock, flags);
6870
6871	if (first)
6872		static_branch_inc(&zones_with_unaccepted_pages);
6873
6874	return true;
6875}
6876
6877#else
6878
6879static bool page_contains_unaccepted(struct page *page, unsigned int order)
6880{
6881	return false;
6882}
6883
6884static void accept_page(struct page *page, unsigned int order)
6885{
6886}
6887
6888static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6889{
6890	return false;
 
 
 
 
 
6891}
6892
6893static inline bool has_unaccepted_memory(void)
6894{
6895	return false;
6896}
6897
6898static bool __free_unaccepted(struct page *page)
6899{
6900	BUILD_BUG();
6901	return false;
6902}
6903
6904#endif /* CONFIG_UNACCEPTED_MEMORY */