2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
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)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node);
78 EXPORT_PER_CPU_SYMBOL(numa_node);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 int _node_numa_mem_[MAX_NUMNODES];
94 * Array of node states.
96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
107 [N_CPU] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page *page)
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
137 page->index = migratetype;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
181 static void __free_pages_ok(struct page *page, unsigned int order);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages);
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names[MIGRATE_TYPES] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor * const compound_page_dtors[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
252 int watermark_scale_factor = 10;
254 static unsigned long __meminitdata nr_kernel_pages;
255 static unsigned long __meminitdata nr_all_pages;
256 static unsigned long __meminitdata dma_reserve;
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
261 static unsigned long __initdata required_kernelcore;
262 static unsigned long __initdata required_movablecore;
263 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
264 static bool mirrored_kernelcore;
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 EXPORT_SYMBOL(movable_zone);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
272 int nr_node_ids __read_mostly = MAX_NUMNODES;
273 int nr_online_nodes __read_mostly = 1;
274 EXPORT_SYMBOL(nr_node_ids);
275 EXPORT_SYMBOL(nr_online_nodes);
278 int page_group_by_mobility_disabled __read_mostly;
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t *pgdat)
283 pgdat->first_deferred_pfn = ULONG_MAX;
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
289 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
297 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
311 unsigned long max_initialise;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
333 static inline void reset_deferred_meminit(pg_data_t *pgdat)
337 static inline bool early_page_uninitialised(unsigned long pfn)
342 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
347 static inline bool update_defer_init(pg_data_t *pgdat,
348 unsigned long pfn, unsigned long zone_end,
349 unsigned long *nr_initialised)
356 void set_pageblock_migratetype(struct page *page, int migratetype)
358 if (unlikely(page_group_by_mobility_disabled &&
359 migratetype < MIGRATE_PCPTYPES))
360 migratetype = MIGRATE_UNMOVABLE;
362 set_pageblock_flags_group(page, (unsigned long)migratetype,
363 PB_migrate, PB_migrate_end);
366 #ifdef CONFIG_DEBUG_VM
367 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
371 unsigned long pfn = page_to_pfn(page);
372 unsigned long sp, start_pfn;
375 seq = zone_span_seqbegin(zone);
376 start_pfn = zone->zone_start_pfn;
377 sp = zone->spanned_pages;
378 if (!zone_spans_pfn(zone, pfn))
380 } while (zone_span_seqretry(zone, seq));
383 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
384 pfn, zone_to_nid(zone), zone->name,
385 start_pfn, start_pfn + sp);
390 static int page_is_consistent(struct zone *zone, struct page *page)
392 if (!pfn_valid_within(page_to_pfn(page)))
394 if (zone != page_zone(page))
400 * Temporary debugging check for pages not lying within a given zone.
402 static int bad_range(struct zone *zone, struct page *page)
404 if (page_outside_zone_boundaries(zone, page))
406 if (!page_is_consistent(zone, page))
412 static inline int bad_range(struct zone *zone, struct page *page)
418 static void bad_page(struct page *page, const char *reason,
419 unsigned long bad_flags)
421 static unsigned long resume;
422 static unsigned long nr_shown;
423 static unsigned long nr_unshown;
425 /* Don't complain about poisoned pages */
426 if (PageHWPoison(page)) {
427 page_mapcount_reset(page); /* remove PageBuddy */
432 * Allow a burst of 60 reports, then keep quiet for that minute;
433 * or allow a steady drip of one report per second.
435 if (nr_shown == 60) {
436 if (time_before(jiffies, resume)) {
442 "BUG: Bad page state: %lu messages suppressed\n",
449 resume = jiffies + 60 * HZ;
451 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
452 current->comm, page_to_pfn(page));
453 __dump_page(page, reason);
454 bad_flags &= page->flags;
456 pr_alert("bad because of flags: %#lx(%pGp)\n",
457 bad_flags, &bad_flags);
458 dump_page_owner(page);
463 /* Leave bad fields for debug, except PageBuddy could make trouble */
464 page_mapcount_reset(page); /* remove PageBuddy */
465 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
469 * Higher-order pages are called "compound pages". They are structured thusly:
471 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
473 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
474 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
476 * The first tail page's ->compound_dtor holds the offset in array of compound
477 * page destructors. See compound_page_dtors.
479 * The first tail page's ->compound_order holds the order of allocation.
480 * This usage means that zero-order pages may not be compound.
483 void free_compound_page(struct page *page)
485 __free_pages_ok(page, compound_order(page));
488 void prep_compound_page(struct page *page, unsigned int order)
491 int nr_pages = 1 << order;
493 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
494 set_compound_order(page, order);
496 for (i = 1; i < nr_pages; i++) {
497 struct page *p = page + i;
498 set_page_count(p, 0);
499 p->mapping = TAIL_MAPPING;
500 set_compound_head(p, page);
502 atomic_set(compound_mapcount_ptr(page), -1);
505 #ifdef CONFIG_DEBUG_PAGEALLOC
506 unsigned int _debug_guardpage_minorder;
507 bool _debug_pagealloc_enabled __read_mostly
508 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
509 EXPORT_SYMBOL(_debug_pagealloc_enabled);
510 bool _debug_guardpage_enabled __read_mostly;
512 static int __init early_debug_pagealloc(char *buf)
517 if (strcmp(buf, "on") == 0)
518 _debug_pagealloc_enabled = true;
520 if (strcmp(buf, "off") == 0)
521 _debug_pagealloc_enabled = false;
525 early_param("debug_pagealloc", early_debug_pagealloc);
527 static bool need_debug_guardpage(void)
529 /* If we don't use debug_pagealloc, we don't need guard page */
530 if (!debug_pagealloc_enabled())
536 static void init_debug_guardpage(void)
538 if (!debug_pagealloc_enabled())
541 _debug_guardpage_enabled = true;
544 struct page_ext_operations debug_guardpage_ops = {
545 .need = need_debug_guardpage,
546 .init = init_debug_guardpage,
549 static int __init debug_guardpage_minorder_setup(char *buf)
553 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
554 pr_err("Bad debug_guardpage_minorder value\n");
557 _debug_guardpage_minorder = res;
558 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
561 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
563 static inline void set_page_guard(struct zone *zone, struct page *page,
564 unsigned int order, int migratetype)
566 struct page_ext *page_ext;
568 if (!debug_guardpage_enabled())
571 page_ext = lookup_page_ext(page);
572 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
574 INIT_LIST_HEAD(&page->lru);
575 set_page_private(page, order);
576 /* Guard pages are not available for any usage */
577 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
580 static inline void clear_page_guard(struct zone *zone, struct page *page,
581 unsigned int order, int migratetype)
583 struct page_ext *page_ext;
585 if (!debug_guardpage_enabled())
588 page_ext = lookup_page_ext(page);
589 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
591 set_page_private(page, 0);
592 if (!is_migrate_isolate(migratetype))
593 __mod_zone_freepage_state(zone, (1 << order), migratetype);
596 struct page_ext_operations debug_guardpage_ops = { NULL, };
597 static inline void set_page_guard(struct zone *zone, struct page *page,
598 unsigned int order, int migratetype) {}
599 static inline void clear_page_guard(struct zone *zone, struct page *page,
600 unsigned int order, int migratetype) {}
603 static inline void set_page_order(struct page *page, unsigned int order)
605 set_page_private(page, order);
606 __SetPageBuddy(page);
609 static inline void rmv_page_order(struct page *page)
611 __ClearPageBuddy(page);
612 set_page_private(page, 0);
616 * This function checks whether a page is free && is the buddy
617 * we can do coalesce a page and its buddy if
618 * (a) the buddy is not in a hole &&
619 * (b) the buddy is in the buddy system &&
620 * (c) a page and its buddy have the same order &&
621 * (d) a page and its buddy are in the same zone.
623 * For recording whether a page is in the buddy system, we set ->_mapcount
624 * PAGE_BUDDY_MAPCOUNT_VALUE.
625 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
626 * serialized by zone->lock.
628 * For recording page's order, we use page_private(page).
630 static inline int page_is_buddy(struct page *page, struct page *buddy,
633 if (!pfn_valid_within(page_to_pfn(buddy)))
636 if (page_is_guard(buddy) && page_order(buddy) == order) {
637 if (page_zone_id(page) != page_zone_id(buddy))
640 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
645 if (PageBuddy(buddy) && page_order(buddy) == order) {
647 * zone check is done late to avoid uselessly
648 * calculating zone/node ids for pages that could
651 if (page_zone_id(page) != page_zone_id(buddy))
654 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
662 * Freeing function for a buddy system allocator.
664 * The concept of a buddy system is to maintain direct-mapped table
665 * (containing bit values) for memory blocks of various "orders".
666 * The bottom level table contains the map for the smallest allocatable
667 * units of memory (here, pages), and each level above it describes
668 * pairs of units from the levels below, hence, "buddies".
669 * At a high level, all that happens here is marking the table entry
670 * at the bottom level available, and propagating the changes upward
671 * as necessary, plus some accounting needed to play nicely with other
672 * parts of the VM system.
673 * At each level, we keep a list of pages, which are heads of continuous
674 * free pages of length of (1 << order) and marked with _mapcount
675 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
677 * So when we are allocating or freeing one, we can derive the state of the
678 * other. That is, if we allocate a small block, and both were
679 * free, the remainder of the region must be split into blocks.
680 * If a block is freed, and its buddy is also free, then this
681 * triggers coalescing into a block of larger size.
686 static inline void __free_one_page(struct page *page,
688 struct zone *zone, unsigned int order,
691 unsigned long page_idx;
692 unsigned long combined_idx;
693 unsigned long uninitialized_var(buddy_idx);
695 unsigned int max_order = MAX_ORDER;
697 VM_BUG_ON(!zone_is_initialized(zone));
698 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
700 VM_BUG_ON(migratetype == -1);
701 if (is_migrate_isolate(migratetype)) {
703 * We restrict max order of merging to prevent merge
704 * between freepages on isolate pageblock and normal
705 * pageblock. Without this, pageblock isolation
706 * could cause incorrect freepage accounting.
708 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
710 __mod_zone_freepage_state(zone, 1 << order, migratetype);
713 page_idx = pfn & ((1 << max_order) - 1);
715 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
716 VM_BUG_ON_PAGE(bad_range(zone, page), page);
718 while (order < max_order - 1) {
719 buddy_idx = __find_buddy_index(page_idx, order);
720 buddy = page + (buddy_idx - page_idx);
721 if (!page_is_buddy(page, buddy, order))
724 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
725 * merge with it and move up one order.
727 if (page_is_guard(buddy)) {
728 clear_page_guard(zone, buddy, order, migratetype);
730 list_del(&buddy->lru);
731 zone->free_area[order].nr_free--;
732 rmv_page_order(buddy);
734 combined_idx = buddy_idx & page_idx;
735 page = page + (combined_idx - page_idx);
736 page_idx = combined_idx;
739 set_page_order(page, order);
742 * If this is not the largest possible page, check if the buddy
743 * of the next-highest order is free. If it is, it's possible
744 * that pages are being freed that will coalesce soon. In case,
745 * that is happening, add the free page to the tail of the list
746 * so it's less likely to be used soon and more likely to be merged
747 * as a higher order page
749 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
750 struct page *higher_page, *higher_buddy;
751 combined_idx = buddy_idx & page_idx;
752 higher_page = page + (combined_idx - page_idx);
753 buddy_idx = __find_buddy_index(combined_idx, order + 1);
754 higher_buddy = higher_page + (buddy_idx - combined_idx);
755 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
756 list_add_tail(&page->lru,
757 &zone->free_area[order].free_list[migratetype]);
762 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
764 zone->free_area[order].nr_free++;
767 static inline int free_pages_check(struct page *page)
769 const char *bad_reason = NULL;
770 unsigned long bad_flags = 0;
772 if (unlikely(atomic_read(&page->_mapcount) != -1))
773 bad_reason = "nonzero mapcount";
774 if (unlikely(page->mapping != NULL))
775 bad_reason = "non-NULL mapping";
776 if (unlikely(page_ref_count(page) != 0))
777 bad_reason = "nonzero _count";
778 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
779 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
780 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
783 if (unlikely(page->mem_cgroup))
784 bad_reason = "page still charged to cgroup";
786 if (unlikely(bad_reason)) {
787 bad_page(page, bad_reason, bad_flags);
790 page_cpupid_reset_last(page);
791 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
792 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
797 * Frees a number of pages from the PCP lists
798 * Assumes all pages on list are in same zone, and of same order.
799 * count is the number of pages to free.
801 * If the zone was previously in an "all pages pinned" state then look to
802 * see if this freeing clears that state.
804 * And clear the zone's pages_scanned counter, to hold off the "all pages are
805 * pinned" detection logic.
807 static void free_pcppages_bulk(struct zone *zone, int count,
808 struct per_cpu_pages *pcp)
813 unsigned long nr_scanned;
815 spin_lock(&zone->lock);
816 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
818 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
822 struct list_head *list;
825 * Remove pages from lists in a round-robin fashion. A
826 * batch_free count is maintained that is incremented when an
827 * empty list is encountered. This is so more pages are freed
828 * off fuller lists instead of spinning excessively around empty
833 if (++migratetype == MIGRATE_PCPTYPES)
835 list = &pcp->lists[migratetype];
836 } while (list_empty(list));
838 /* This is the only non-empty list. Free them all. */
839 if (batch_free == MIGRATE_PCPTYPES)
840 batch_free = to_free;
843 int mt; /* migratetype of the to-be-freed page */
845 page = list_last_entry(list, struct page, lru);
846 /* must delete as __free_one_page list manipulates */
847 list_del(&page->lru);
849 mt = get_pcppage_migratetype(page);
850 /* MIGRATE_ISOLATE page should not go to pcplists */
851 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
852 /* Pageblock could have been isolated meanwhile */
853 if (unlikely(has_isolate_pageblock(zone)))
854 mt = get_pageblock_migratetype(page);
856 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
857 trace_mm_page_pcpu_drain(page, 0, mt);
858 } while (--to_free && --batch_free && !list_empty(list));
860 spin_unlock(&zone->lock);
863 static void free_one_page(struct zone *zone,
864 struct page *page, unsigned long pfn,
868 unsigned long nr_scanned;
869 spin_lock(&zone->lock);
870 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
872 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
874 if (unlikely(has_isolate_pageblock(zone) ||
875 is_migrate_isolate(migratetype))) {
876 migratetype = get_pfnblock_migratetype(page, pfn);
878 __free_one_page(page, pfn, zone, order, migratetype);
879 spin_unlock(&zone->lock);
882 static int free_tail_pages_check(struct page *head_page, struct page *page)
887 * We rely page->lru.next never has bit 0 set, unless the page
888 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
890 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
892 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
896 switch (page - head_page) {
898 /* the first tail page: ->mapping is compound_mapcount() */
899 if (unlikely(compound_mapcount(page))) {
900 bad_page(page, "nonzero compound_mapcount", 0);
906 * the second tail page: ->mapping is
907 * page_deferred_list().next -- ignore value.
911 if (page->mapping != TAIL_MAPPING) {
912 bad_page(page, "corrupted mapping in tail page", 0);
917 if (unlikely(!PageTail(page))) {
918 bad_page(page, "PageTail not set", 0);
921 if (unlikely(compound_head(page) != head_page)) {
922 bad_page(page, "compound_head not consistent", 0);
927 page->mapping = NULL;
928 clear_compound_head(page);
932 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
933 unsigned long zone, int nid)
935 set_page_links(page, zone, nid, pfn);
936 init_page_count(page);
937 page_mapcount_reset(page);
938 page_cpupid_reset_last(page);
940 INIT_LIST_HEAD(&page->lru);
941 #ifdef WANT_PAGE_VIRTUAL
942 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
943 if (!is_highmem_idx(zone))
944 set_page_address(page, __va(pfn << PAGE_SHIFT));
948 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
951 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
954 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
955 static void init_reserved_page(unsigned long pfn)
960 if (!early_page_uninitialised(pfn))
963 nid = early_pfn_to_nid(pfn);
964 pgdat = NODE_DATA(nid);
966 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
967 struct zone *zone = &pgdat->node_zones[zid];
969 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
972 __init_single_pfn(pfn, zid, nid);
975 static inline void init_reserved_page(unsigned long pfn)
978 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
981 * Initialised pages do not have PageReserved set. This function is
982 * called for each range allocated by the bootmem allocator and
983 * marks the pages PageReserved. The remaining valid pages are later
984 * sent to the buddy page allocator.
986 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
988 unsigned long start_pfn = PFN_DOWN(start);
989 unsigned long end_pfn = PFN_UP(end);
991 for (; start_pfn < end_pfn; start_pfn++) {
992 if (pfn_valid(start_pfn)) {
993 struct page *page = pfn_to_page(start_pfn);
995 init_reserved_page(start_pfn);
997 /* Avoid false-positive PageTail() */
998 INIT_LIST_HEAD(&page->lru);
1000 SetPageReserved(page);
1005 static bool free_pages_prepare(struct page *page, unsigned int order)
1007 bool compound = PageCompound(page);
1010 VM_BUG_ON_PAGE(PageTail(page), page);
1011 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1013 trace_mm_page_free(page, order);
1014 kmemcheck_free_shadow(page, order);
1015 kasan_free_pages(page, order);
1018 page->mapping = NULL;
1019 bad += free_pages_check(page);
1020 for (i = 1; i < (1 << order); i++) {
1022 bad += free_tail_pages_check(page, page + i);
1023 bad += free_pages_check(page + i);
1028 reset_page_owner(page, order);
1030 if (!PageHighMem(page)) {
1031 debug_check_no_locks_freed(page_address(page),
1032 PAGE_SIZE << order);
1033 debug_check_no_obj_freed(page_address(page),
1034 PAGE_SIZE << order);
1036 arch_free_page(page, order);
1037 kernel_poison_pages(page, 1 << order, 0);
1038 kernel_map_pages(page, 1 << order, 0);
1043 static void __free_pages_ok(struct page *page, unsigned int order)
1045 unsigned long flags;
1047 unsigned long pfn = page_to_pfn(page);
1049 if (!free_pages_prepare(page, order))
1052 migratetype = get_pfnblock_migratetype(page, pfn);
1053 local_irq_save(flags);
1054 __count_vm_events(PGFREE, 1 << order);
1055 free_one_page(page_zone(page), page, pfn, order, migratetype);
1056 local_irq_restore(flags);
1059 static void __init __free_pages_boot_core(struct page *page,
1060 unsigned long pfn, unsigned int order)
1062 unsigned int nr_pages = 1 << order;
1063 struct page *p = page;
1067 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1069 __ClearPageReserved(p);
1070 set_page_count(p, 0);
1072 __ClearPageReserved(p);
1073 set_page_count(p, 0);
1075 page_zone(page)->managed_pages += nr_pages;
1076 set_page_refcounted(page);
1077 __free_pages(page, order);
1080 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1081 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1083 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1085 int __meminit early_pfn_to_nid(unsigned long pfn)
1087 static DEFINE_SPINLOCK(early_pfn_lock);
1090 spin_lock(&early_pfn_lock);
1091 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1094 spin_unlock(&early_pfn_lock);
1100 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1101 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1102 struct mminit_pfnnid_cache *state)
1106 nid = __early_pfn_to_nid(pfn, state);
1107 if (nid >= 0 && nid != node)
1112 /* Only safe to use early in boot when initialisation is single-threaded */
1113 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1115 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1120 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1124 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1125 struct mminit_pfnnid_cache *state)
1132 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1135 if (early_page_uninitialised(pfn))
1137 return __free_pages_boot_core(page, pfn, order);
1141 * Check that the whole (or subset of) a pageblock given by the interval of
1142 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1143 * with the migration of free compaction scanner. The scanners then need to
1144 * use only pfn_valid_within() check for arches that allow holes within
1147 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1149 * It's possible on some configurations to have a setup like node0 node1 node0
1150 * i.e. it's possible that all pages within a zones range of pages do not
1151 * belong to a single zone. We assume that a border between node0 and node1
1152 * can occur within a single pageblock, but not a node0 node1 node0
1153 * interleaving within a single pageblock. It is therefore sufficient to check
1154 * the first and last page of a pageblock and avoid checking each individual
1155 * page in a pageblock.
1157 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1158 unsigned long end_pfn, struct zone *zone)
1160 struct page *start_page;
1161 struct page *end_page;
1163 /* end_pfn is one past the range we are checking */
1166 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1169 start_page = pfn_to_page(start_pfn);
1171 if (page_zone(start_page) != zone)
1174 end_page = pfn_to_page(end_pfn);
1176 /* This gives a shorter code than deriving page_zone(end_page) */
1177 if (page_zone_id(start_page) != page_zone_id(end_page))
1183 void set_zone_contiguous(struct zone *zone)
1185 unsigned long block_start_pfn = zone->zone_start_pfn;
1186 unsigned long block_end_pfn;
1188 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1189 for (; block_start_pfn < zone_end_pfn(zone);
1190 block_start_pfn = block_end_pfn,
1191 block_end_pfn += pageblock_nr_pages) {
1193 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1195 if (!__pageblock_pfn_to_page(block_start_pfn,
1196 block_end_pfn, zone))
1200 /* We confirm that there is no hole */
1201 zone->contiguous = true;
1204 void clear_zone_contiguous(struct zone *zone)
1206 zone->contiguous = false;
1209 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1210 static void __init deferred_free_range(struct page *page,
1211 unsigned long pfn, int nr_pages)
1218 /* Free a large naturally-aligned chunk if possible */
1219 if (nr_pages == MAX_ORDER_NR_PAGES &&
1220 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1221 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1222 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1226 for (i = 0; i < nr_pages; i++, page++, pfn++)
1227 __free_pages_boot_core(page, pfn, 0);
1230 /* Completion tracking for deferred_init_memmap() threads */
1231 static atomic_t pgdat_init_n_undone __initdata;
1232 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1234 static inline void __init pgdat_init_report_one_done(void)
1236 if (atomic_dec_and_test(&pgdat_init_n_undone))
1237 complete(&pgdat_init_all_done_comp);
1240 /* Initialise remaining memory on a node */
1241 static int __init deferred_init_memmap(void *data)
1243 pg_data_t *pgdat = data;
1244 int nid = pgdat->node_id;
1245 struct mminit_pfnnid_cache nid_init_state = { };
1246 unsigned long start = jiffies;
1247 unsigned long nr_pages = 0;
1248 unsigned long walk_start, walk_end;
1251 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1252 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1254 if (first_init_pfn == ULONG_MAX) {
1255 pgdat_init_report_one_done();
1259 /* Bind memory initialisation thread to a local node if possible */
1260 if (!cpumask_empty(cpumask))
1261 set_cpus_allowed_ptr(current, cpumask);
1263 /* Sanity check boundaries */
1264 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1265 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1266 pgdat->first_deferred_pfn = ULONG_MAX;
1268 /* Only the highest zone is deferred so find it */
1269 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1270 zone = pgdat->node_zones + zid;
1271 if (first_init_pfn < zone_end_pfn(zone))
1275 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1276 unsigned long pfn, end_pfn;
1277 struct page *page = NULL;
1278 struct page *free_base_page = NULL;
1279 unsigned long free_base_pfn = 0;
1282 end_pfn = min(walk_end, zone_end_pfn(zone));
1283 pfn = first_init_pfn;
1284 if (pfn < walk_start)
1286 if (pfn < zone->zone_start_pfn)
1287 pfn = zone->zone_start_pfn;
1289 for (; pfn < end_pfn; pfn++) {
1290 if (!pfn_valid_within(pfn))
1294 * Ensure pfn_valid is checked every
1295 * MAX_ORDER_NR_PAGES for memory holes
1297 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1298 if (!pfn_valid(pfn)) {
1304 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1309 /* Minimise pfn page lookups and scheduler checks */
1310 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1313 nr_pages += nr_to_free;
1314 deferred_free_range(free_base_page,
1315 free_base_pfn, nr_to_free);
1316 free_base_page = NULL;
1317 free_base_pfn = nr_to_free = 0;
1319 page = pfn_to_page(pfn);
1324 VM_BUG_ON(page_zone(page) != zone);
1328 __init_single_page(page, pfn, zid, nid);
1329 if (!free_base_page) {
1330 free_base_page = page;
1331 free_base_pfn = pfn;
1336 /* Where possible, batch up pages for a single free */
1339 /* Free the current block of pages to allocator */
1340 nr_pages += nr_to_free;
1341 deferred_free_range(free_base_page, free_base_pfn,
1343 free_base_page = NULL;
1344 free_base_pfn = nr_to_free = 0;
1347 first_init_pfn = max(end_pfn, first_init_pfn);
1350 /* Sanity check that the next zone really is unpopulated */
1351 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1353 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1354 jiffies_to_msecs(jiffies - start));
1356 pgdat_init_report_one_done();
1359 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1361 void __init page_alloc_init_late(void)
1365 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1368 /* There will be num_node_state(N_MEMORY) threads */
1369 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1370 for_each_node_state(nid, N_MEMORY) {
1371 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1374 /* Block until all are initialised */
1375 wait_for_completion(&pgdat_init_all_done_comp);
1377 /* Reinit limits that are based on free pages after the kernel is up */
1378 files_maxfiles_init();
1381 for_each_populated_zone(zone)
1382 set_zone_contiguous(zone);
1386 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1387 void __init init_cma_reserved_pageblock(struct page *page)
1389 unsigned i = pageblock_nr_pages;
1390 struct page *p = page;
1393 __ClearPageReserved(p);
1394 set_page_count(p, 0);
1397 set_pageblock_migratetype(page, MIGRATE_CMA);
1399 if (pageblock_order >= MAX_ORDER) {
1400 i = pageblock_nr_pages;
1403 set_page_refcounted(p);
1404 __free_pages(p, MAX_ORDER - 1);
1405 p += MAX_ORDER_NR_PAGES;
1406 } while (i -= MAX_ORDER_NR_PAGES);
1408 set_page_refcounted(page);
1409 __free_pages(page, pageblock_order);
1412 adjust_managed_page_count(page, pageblock_nr_pages);
1417 * The order of subdivision here is critical for the IO subsystem.
1418 * Please do not alter this order without good reasons and regression
1419 * testing. Specifically, as large blocks of memory are subdivided,
1420 * the order in which smaller blocks are delivered depends on the order
1421 * they're subdivided in this function. This is the primary factor
1422 * influencing the order in which pages are delivered to the IO
1423 * subsystem according to empirical testing, and this is also justified
1424 * by considering the behavior of a buddy system containing a single
1425 * large block of memory acted on by a series of small allocations.
1426 * This behavior is a critical factor in sglist merging's success.
1430 static inline void expand(struct zone *zone, struct page *page,
1431 int low, int high, struct free_area *area,
1434 unsigned long size = 1 << high;
1436 while (high > low) {
1440 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1442 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1443 debug_guardpage_enabled() &&
1444 high < debug_guardpage_minorder()) {
1446 * Mark as guard pages (or page), that will allow to
1447 * merge back to allocator when buddy will be freed.
1448 * Corresponding page table entries will not be touched,
1449 * pages will stay not present in virtual address space
1451 set_page_guard(zone, &page[size], high, migratetype);
1454 list_add(&page[size].lru, &area->free_list[migratetype]);
1456 set_page_order(&page[size], high);
1461 * This page is about to be returned from the page allocator
1463 static inline int check_new_page(struct page *page)
1465 const char *bad_reason = NULL;
1466 unsigned long bad_flags = 0;
1468 if (unlikely(atomic_read(&page->_mapcount) != -1))
1469 bad_reason = "nonzero mapcount";
1470 if (unlikely(page->mapping != NULL))
1471 bad_reason = "non-NULL mapping";
1472 if (unlikely(page_ref_count(page) != 0))
1473 bad_reason = "nonzero _count";
1474 if (unlikely(page->flags & __PG_HWPOISON)) {
1475 bad_reason = "HWPoisoned (hardware-corrupted)";
1476 bad_flags = __PG_HWPOISON;
1478 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1479 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1480 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1483 if (unlikely(page->mem_cgroup))
1484 bad_reason = "page still charged to cgroup";
1486 if (unlikely(bad_reason)) {
1487 bad_page(page, bad_reason, bad_flags);
1493 static inline bool free_pages_prezeroed(bool poisoned)
1495 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1496 page_poisoning_enabled() && poisoned;
1499 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1503 bool poisoned = true;
1505 for (i = 0; i < (1 << order); i++) {
1506 struct page *p = page + i;
1507 if (unlikely(check_new_page(p)))
1510 poisoned &= page_is_poisoned(p);
1513 set_page_private(page, 0);
1514 set_page_refcounted(page);
1516 arch_alloc_page(page, order);
1517 kernel_map_pages(page, 1 << order, 1);
1518 kernel_poison_pages(page, 1 << order, 1);
1519 kasan_alloc_pages(page, order);
1521 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1522 for (i = 0; i < (1 << order); i++)
1523 clear_highpage(page + i);
1525 if (order && (gfp_flags & __GFP_COMP))
1526 prep_compound_page(page, order);
1528 set_page_owner(page, order, gfp_flags);
1531 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1532 * allocate the page. The expectation is that the caller is taking
1533 * steps that will free more memory. The caller should avoid the page
1534 * being used for !PFMEMALLOC purposes.
1536 if (alloc_flags & ALLOC_NO_WATERMARKS)
1537 set_page_pfmemalloc(page);
1539 clear_page_pfmemalloc(page);
1545 * Go through the free lists for the given migratetype and remove
1546 * the smallest available page from the freelists
1549 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1552 unsigned int current_order;
1553 struct free_area *area;
1556 /* Find a page of the appropriate size in the preferred list */
1557 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1558 area = &(zone->free_area[current_order]);
1559 page = list_first_entry_or_null(&area->free_list[migratetype],
1563 list_del(&page->lru);
1564 rmv_page_order(page);
1566 expand(zone, page, order, current_order, area, migratetype);
1567 set_pcppage_migratetype(page, migratetype);
1576 * This array describes the order lists are fallen back to when
1577 * the free lists for the desirable migrate type are depleted
1579 static int fallbacks[MIGRATE_TYPES][4] = {
1580 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1581 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1582 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1584 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1586 #ifdef CONFIG_MEMORY_ISOLATION
1587 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1592 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1595 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1598 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1599 unsigned int order) { return NULL; }
1603 * Move the free pages in a range to the free lists of the requested type.
1604 * Note that start_page and end_pages are not aligned on a pageblock
1605 * boundary. If alignment is required, use move_freepages_block()
1607 int move_freepages(struct zone *zone,
1608 struct page *start_page, struct page *end_page,
1613 int pages_moved = 0;
1615 #ifndef CONFIG_HOLES_IN_ZONE
1617 * page_zone is not safe to call in this context when
1618 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1619 * anyway as we check zone boundaries in move_freepages_block().
1620 * Remove at a later date when no bug reports exist related to
1621 * grouping pages by mobility
1623 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1626 for (page = start_page; page <= end_page;) {
1627 /* Make sure we are not inadvertently changing nodes */
1628 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1630 if (!pfn_valid_within(page_to_pfn(page))) {
1635 if (!PageBuddy(page)) {
1640 order = page_order(page);
1641 list_move(&page->lru,
1642 &zone->free_area[order].free_list[migratetype]);
1644 pages_moved += 1 << order;
1650 int move_freepages_block(struct zone *zone, struct page *page,
1653 unsigned long start_pfn, end_pfn;
1654 struct page *start_page, *end_page;
1656 start_pfn = page_to_pfn(page);
1657 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1658 start_page = pfn_to_page(start_pfn);
1659 end_page = start_page + pageblock_nr_pages - 1;
1660 end_pfn = start_pfn + pageblock_nr_pages - 1;
1662 /* Do not cross zone boundaries */
1663 if (!zone_spans_pfn(zone, start_pfn))
1665 if (!zone_spans_pfn(zone, end_pfn))
1668 return move_freepages(zone, start_page, end_page, migratetype);
1671 static void change_pageblock_range(struct page *pageblock_page,
1672 int start_order, int migratetype)
1674 int nr_pageblocks = 1 << (start_order - pageblock_order);
1676 while (nr_pageblocks--) {
1677 set_pageblock_migratetype(pageblock_page, migratetype);
1678 pageblock_page += pageblock_nr_pages;
1683 * When we are falling back to another migratetype during allocation, try to
1684 * steal extra free pages from the same pageblocks to satisfy further
1685 * allocations, instead of polluting multiple pageblocks.
1687 * If we are stealing a relatively large buddy page, it is likely there will
1688 * be more free pages in the pageblock, so try to steal them all. For
1689 * reclaimable and unmovable allocations, we steal regardless of page size,
1690 * as fragmentation caused by those allocations polluting movable pageblocks
1691 * is worse than movable allocations stealing from unmovable and reclaimable
1694 static bool can_steal_fallback(unsigned int order, int start_mt)
1697 * Leaving this order check is intended, although there is
1698 * relaxed order check in next check. The reason is that
1699 * we can actually steal whole pageblock if this condition met,
1700 * but, below check doesn't guarantee it and that is just heuristic
1701 * so could be changed anytime.
1703 if (order >= pageblock_order)
1706 if (order >= pageblock_order / 2 ||
1707 start_mt == MIGRATE_RECLAIMABLE ||
1708 start_mt == MIGRATE_UNMOVABLE ||
1709 page_group_by_mobility_disabled)
1716 * This function implements actual steal behaviour. If order is large enough,
1717 * we can steal whole pageblock. If not, we first move freepages in this
1718 * pageblock and check whether half of pages are moved or not. If half of
1719 * pages are moved, we can change migratetype of pageblock and permanently
1720 * use it's pages as requested migratetype in the future.
1722 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1725 unsigned int current_order = page_order(page);
1728 /* Take ownership for orders >= pageblock_order */
1729 if (current_order >= pageblock_order) {
1730 change_pageblock_range(page, current_order, start_type);
1734 pages = move_freepages_block(zone, page, start_type);
1736 /* Claim the whole block if over half of it is free */
1737 if (pages >= (1 << (pageblock_order-1)) ||
1738 page_group_by_mobility_disabled)
1739 set_pageblock_migratetype(page, start_type);
1743 * Check whether there is a suitable fallback freepage with requested order.
1744 * If only_stealable is true, this function returns fallback_mt only if
1745 * we can steal other freepages all together. This would help to reduce
1746 * fragmentation due to mixed migratetype pages in one pageblock.
1748 int find_suitable_fallback(struct free_area *area, unsigned int order,
1749 int migratetype, bool only_stealable, bool *can_steal)
1754 if (area->nr_free == 0)
1759 fallback_mt = fallbacks[migratetype][i];
1760 if (fallback_mt == MIGRATE_TYPES)
1763 if (list_empty(&area->free_list[fallback_mt]))
1766 if (can_steal_fallback(order, migratetype))
1769 if (!only_stealable)
1780 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1781 * there are no empty page blocks that contain a page with a suitable order
1783 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1784 unsigned int alloc_order)
1787 unsigned long max_managed, flags;
1790 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1791 * Check is race-prone but harmless.
1793 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1794 if (zone->nr_reserved_highatomic >= max_managed)
1797 spin_lock_irqsave(&zone->lock, flags);
1799 /* Recheck the nr_reserved_highatomic limit under the lock */
1800 if (zone->nr_reserved_highatomic >= max_managed)
1804 mt = get_pageblock_migratetype(page);
1805 if (mt != MIGRATE_HIGHATOMIC &&
1806 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1807 zone->nr_reserved_highatomic += pageblock_nr_pages;
1808 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1809 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1813 spin_unlock_irqrestore(&zone->lock, flags);
1817 * Used when an allocation is about to fail under memory pressure. This
1818 * potentially hurts the reliability of high-order allocations when under
1819 * intense memory pressure but failed atomic allocations should be easier
1820 * to recover from than an OOM.
1822 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1824 struct zonelist *zonelist = ac->zonelist;
1825 unsigned long flags;
1831 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1833 /* Preserve at least one pageblock */
1834 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1837 spin_lock_irqsave(&zone->lock, flags);
1838 for (order = 0; order < MAX_ORDER; order++) {
1839 struct free_area *area = &(zone->free_area[order]);
1841 page = list_first_entry_or_null(
1842 &area->free_list[MIGRATE_HIGHATOMIC],
1848 * It should never happen but changes to locking could
1849 * inadvertently allow a per-cpu drain to add pages
1850 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1851 * and watch for underflows.
1853 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1854 zone->nr_reserved_highatomic);
1857 * Convert to ac->migratetype and avoid the normal
1858 * pageblock stealing heuristics. Minimally, the caller
1859 * is doing the work and needs the pages. More
1860 * importantly, if the block was always converted to
1861 * MIGRATE_UNMOVABLE or another type then the number
1862 * of pageblocks that cannot be completely freed
1865 set_pageblock_migratetype(page, ac->migratetype);
1866 move_freepages_block(zone, page, ac->migratetype);
1867 spin_unlock_irqrestore(&zone->lock, flags);
1870 spin_unlock_irqrestore(&zone->lock, flags);
1874 /* Remove an element from the buddy allocator from the fallback list */
1875 static inline struct page *
1876 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1878 struct free_area *area;
1879 unsigned int current_order;
1884 /* Find the largest possible block of pages in the other list */
1885 for (current_order = MAX_ORDER-1;
1886 current_order >= order && current_order <= MAX_ORDER-1;
1888 area = &(zone->free_area[current_order]);
1889 fallback_mt = find_suitable_fallback(area, current_order,
1890 start_migratetype, false, &can_steal);
1891 if (fallback_mt == -1)
1894 page = list_first_entry(&area->free_list[fallback_mt],
1897 steal_suitable_fallback(zone, page, start_migratetype);
1899 /* Remove the page from the freelists */
1901 list_del(&page->lru);
1902 rmv_page_order(page);
1904 expand(zone, page, order, current_order, area,
1907 * The pcppage_migratetype may differ from pageblock's
1908 * migratetype depending on the decisions in
1909 * find_suitable_fallback(). This is OK as long as it does not
1910 * differ for MIGRATE_CMA pageblocks. Those can be used as
1911 * fallback only via special __rmqueue_cma_fallback() function
1913 set_pcppage_migratetype(page, start_migratetype);
1915 trace_mm_page_alloc_extfrag(page, order, current_order,
1916 start_migratetype, fallback_mt);
1925 * Do the hard work of removing an element from the buddy allocator.
1926 * Call me with the zone->lock already held.
1928 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1933 page = __rmqueue_smallest(zone, order, migratetype);
1934 if (unlikely(!page)) {
1935 if (migratetype == MIGRATE_MOVABLE)
1936 page = __rmqueue_cma_fallback(zone, order);
1939 page = __rmqueue_fallback(zone, order, migratetype);
1942 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1947 * Obtain a specified number of elements from the buddy allocator, all under
1948 * a single hold of the lock, for efficiency. Add them to the supplied list.
1949 * Returns the number of new pages which were placed at *list.
1951 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1952 unsigned long count, struct list_head *list,
1953 int migratetype, bool cold)
1957 spin_lock(&zone->lock);
1958 for (i = 0; i < count; ++i) {
1959 struct page *page = __rmqueue(zone, order, migratetype);
1960 if (unlikely(page == NULL))
1964 * Split buddy pages returned by expand() are received here
1965 * in physical page order. The page is added to the callers and
1966 * list and the list head then moves forward. From the callers
1967 * perspective, the linked list is ordered by page number in
1968 * some conditions. This is useful for IO devices that can
1969 * merge IO requests if the physical pages are ordered
1973 list_add(&page->lru, list);
1975 list_add_tail(&page->lru, list);
1977 if (is_migrate_cma(get_pcppage_migratetype(page)))
1978 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1981 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1982 spin_unlock(&zone->lock);
1988 * Called from the vmstat counter updater to drain pagesets of this
1989 * currently executing processor on remote nodes after they have
1992 * Note that this function must be called with the thread pinned to
1993 * a single processor.
1995 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1997 unsigned long flags;
1998 int to_drain, batch;
2000 local_irq_save(flags);
2001 batch = READ_ONCE(pcp->batch);
2002 to_drain = min(pcp->count, batch);
2004 free_pcppages_bulk(zone, to_drain, pcp);
2005 pcp->count -= to_drain;
2007 local_irq_restore(flags);
2012 * Drain pcplists of the indicated processor and zone.
2014 * The processor must either be the current processor and the
2015 * thread pinned to the current processor or a processor that
2018 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2020 unsigned long flags;
2021 struct per_cpu_pageset *pset;
2022 struct per_cpu_pages *pcp;
2024 local_irq_save(flags);
2025 pset = per_cpu_ptr(zone->pageset, cpu);
2029 free_pcppages_bulk(zone, pcp->count, pcp);
2032 local_irq_restore(flags);
2036 * Drain pcplists of all zones on the indicated processor.
2038 * The processor must either be the current processor and the
2039 * thread pinned to the current processor or a processor that
2042 static void drain_pages(unsigned int cpu)
2046 for_each_populated_zone(zone) {
2047 drain_pages_zone(cpu, zone);
2052 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2054 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2055 * the single zone's pages.
2057 void drain_local_pages(struct zone *zone)
2059 int cpu = smp_processor_id();
2062 drain_pages_zone(cpu, zone);
2068 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2070 * When zone parameter is non-NULL, spill just the single zone's pages.
2072 * Note that this code is protected against sending an IPI to an offline
2073 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2074 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2075 * nothing keeps CPUs from showing up after we populated the cpumask and
2076 * before the call to on_each_cpu_mask().
2078 void drain_all_pages(struct zone *zone)
2083 * Allocate in the BSS so we wont require allocation in
2084 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2086 static cpumask_t cpus_with_pcps;
2089 * We don't care about racing with CPU hotplug event
2090 * as offline notification will cause the notified
2091 * cpu to drain that CPU pcps and on_each_cpu_mask
2092 * disables preemption as part of its processing
2094 for_each_online_cpu(cpu) {
2095 struct per_cpu_pageset *pcp;
2097 bool has_pcps = false;
2100 pcp = per_cpu_ptr(zone->pageset, cpu);
2104 for_each_populated_zone(z) {
2105 pcp = per_cpu_ptr(z->pageset, cpu);
2106 if (pcp->pcp.count) {
2114 cpumask_set_cpu(cpu, &cpus_with_pcps);
2116 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2118 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2122 #ifdef CONFIG_HIBERNATION
2124 void mark_free_pages(struct zone *zone)
2126 unsigned long pfn, max_zone_pfn;
2127 unsigned long flags;
2128 unsigned int order, t;
2131 if (zone_is_empty(zone))
2134 spin_lock_irqsave(&zone->lock, flags);
2136 max_zone_pfn = zone_end_pfn(zone);
2137 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2138 if (pfn_valid(pfn)) {
2139 page = pfn_to_page(pfn);
2140 if (!swsusp_page_is_forbidden(page))
2141 swsusp_unset_page_free(page);
2144 for_each_migratetype_order(order, t) {
2145 list_for_each_entry(page,
2146 &zone->free_area[order].free_list[t], lru) {
2149 pfn = page_to_pfn(page);
2150 for (i = 0; i < (1UL << order); i++)
2151 swsusp_set_page_free(pfn_to_page(pfn + i));
2154 spin_unlock_irqrestore(&zone->lock, flags);
2156 #endif /* CONFIG_PM */
2159 * Free a 0-order page
2160 * cold == true ? free a cold page : free a hot page
2162 void free_hot_cold_page(struct page *page, bool cold)
2164 struct zone *zone = page_zone(page);
2165 struct per_cpu_pages *pcp;
2166 unsigned long flags;
2167 unsigned long pfn = page_to_pfn(page);
2170 if (!free_pages_prepare(page, 0))
2173 migratetype = get_pfnblock_migratetype(page, pfn);
2174 set_pcppage_migratetype(page, migratetype);
2175 local_irq_save(flags);
2176 __count_vm_event(PGFREE);
2179 * We only track unmovable, reclaimable and movable on pcp lists.
2180 * Free ISOLATE pages back to the allocator because they are being
2181 * offlined but treat RESERVE as movable pages so we can get those
2182 * areas back if necessary. Otherwise, we may have to free
2183 * excessively into the page allocator
2185 if (migratetype >= MIGRATE_PCPTYPES) {
2186 if (unlikely(is_migrate_isolate(migratetype))) {
2187 free_one_page(zone, page, pfn, 0, migratetype);
2190 migratetype = MIGRATE_MOVABLE;
2193 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2195 list_add(&page->lru, &pcp->lists[migratetype]);
2197 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2199 if (pcp->count >= pcp->high) {
2200 unsigned long batch = READ_ONCE(pcp->batch);
2201 free_pcppages_bulk(zone, batch, pcp);
2202 pcp->count -= batch;
2206 local_irq_restore(flags);
2210 * Free a list of 0-order pages
2212 void free_hot_cold_page_list(struct list_head *list, bool cold)
2214 struct page *page, *next;
2216 list_for_each_entry_safe(page, next, list, lru) {
2217 trace_mm_page_free_batched(page, cold);
2218 free_hot_cold_page(page, cold);
2223 * split_page takes a non-compound higher-order page, and splits it into
2224 * n (1<<order) sub-pages: page[0..n]
2225 * Each sub-page must be freed individually.
2227 * Note: this is probably too low level an operation for use in drivers.
2228 * Please consult with lkml before using this in your driver.
2230 void split_page(struct page *page, unsigned int order)
2235 VM_BUG_ON_PAGE(PageCompound(page), page);
2236 VM_BUG_ON_PAGE(!page_count(page), page);
2238 #ifdef CONFIG_KMEMCHECK
2240 * Split shadow pages too, because free(page[0]) would
2241 * otherwise free the whole shadow.
2243 if (kmemcheck_page_is_tracked(page))
2244 split_page(virt_to_page(page[0].shadow), order);
2247 gfp_mask = get_page_owner_gfp(page);
2248 set_page_owner(page, 0, gfp_mask);
2249 for (i = 1; i < (1 << order); i++) {
2250 set_page_refcounted(page + i);
2251 set_page_owner(page + i, 0, gfp_mask);
2254 EXPORT_SYMBOL_GPL(split_page);
2256 int __isolate_free_page(struct page *page, unsigned int order)
2258 unsigned long watermark;
2262 BUG_ON(!PageBuddy(page));
2264 zone = page_zone(page);
2265 mt = get_pageblock_migratetype(page);
2267 if (!is_migrate_isolate(mt)) {
2268 /* Obey watermarks as if the page was being allocated */
2269 watermark = low_wmark_pages(zone) + (1 << order);
2270 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2273 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2276 /* Remove page from free list */
2277 list_del(&page->lru);
2278 zone->free_area[order].nr_free--;
2279 rmv_page_order(page);
2281 set_page_owner(page, order, __GFP_MOVABLE);
2283 /* Set the pageblock if the isolated page is at least a pageblock */
2284 if (order >= pageblock_order - 1) {
2285 struct page *endpage = page + (1 << order) - 1;
2286 for (; page < endpage; page += pageblock_nr_pages) {
2287 int mt = get_pageblock_migratetype(page);
2288 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2289 set_pageblock_migratetype(page,
2295 return 1UL << order;
2299 * Similar to split_page except the page is already free. As this is only
2300 * being used for migration, the migratetype of the block also changes.
2301 * As this is called with interrupts disabled, the caller is responsible
2302 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2305 * Note: this is probably too low level an operation for use in drivers.
2306 * Please consult with lkml before using this in your driver.
2308 int split_free_page(struct page *page)
2313 order = page_order(page);
2315 nr_pages = __isolate_free_page(page, order);
2319 /* Split into individual pages */
2320 set_page_refcounted(page);
2321 split_page(page, order);
2326 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2329 struct page *buffered_rmqueue(struct zone *preferred_zone,
2330 struct zone *zone, unsigned int order,
2331 gfp_t gfp_flags, int alloc_flags, int migratetype)
2333 unsigned long flags;
2335 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2337 if (likely(order == 0)) {
2338 struct per_cpu_pages *pcp;
2339 struct list_head *list;
2341 local_irq_save(flags);
2342 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2343 list = &pcp->lists[migratetype];
2344 if (list_empty(list)) {
2345 pcp->count += rmqueue_bulk(zone, 0,
2348 if (unlikely(list_empty(list)))
2353 page = list_last_entry(list, struct page, lru);
2355 page = list_first_entry(list, struct page, lru);
2357 list_del(&page->lru);
2361 * We most definitely don't want callers attempting to
2362 * allocate greater than order-1 page units with __GFP_NOFAIL.
2364 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2365 spin_lock_irqsave(&zone->lock, flags);
2368 if (alloc_flags & ALLOC_HARDER) {
2369 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2371 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2374 page = __rmqueue(zone, order, migratetype);
2375 spin_unlock(&zone->lock);
2378 __mod_zone_freepage_state(zone, -(1 << order),
2379 get_pcppage_migratetype(page));
2382 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2383 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2384 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2385 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2387 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2388 zone_statistics(preferred_zone, zone, gfp_flags);
2389 local_irq_restore(flags);
2391 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2395 local_irq_restore(flags);
2399 #ifdef CONFIG_FAIL_PAGE_ALLOC
2402 struct fault_attr attr;
2404 bool ignore_gfp_highmem;
2405 bool ignore_gfp_reclaim;
2407 } fail_page_alloc = {
2408 .attr = FAULT_ATTR_INITIALIZER,
2409 .ignore_gfp_reclaim = true,
2410 .ignore_gfp_highmem = true,
2414 static int __init setup_fail_page_alloc(char *str)
2416 return setup_fault_attr(&fail_page_alloc.attr, str);
2418 __setup("fail_page_alloc=", setup_fail_page_alloc);
2420 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2422 if (order < fail_page_alloc.min_order)
2424 if (gfp_mask & __GFP_NOFAIL)
2426 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2428 if (fail_page_alloc.ignore_gfp_reclaim &&
2429 (gfp_mask & __GFP_DIRECT_RECLAIM))
2432 return should_fail(&fail_page_alloc.attr, 1 << order);
2435 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2437 static int __init fail_page_alloc_debugfs(void)
2439 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2442 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2443 &fail_page_alloc.attr);
2445 return PTR_ERR(dir);
2447 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2448 &fail_page_alloc.ignore_gfp_reclaim))
2450 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2451 &fail_page_alloc.ignore_gfp_highmem))
2453 if (!debugfs_create_u32("min-order", mode, dir,
2454 &fail_page_alloc.min_order))
2459 debugfs_remove_recursive(dir);
2464 late_initcall(fail_page_alloc_debugfs);
2466 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2468 #else /* CONFIG_FAIL_PAGE_ALLOC */
2470 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2475 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2478 * Return true if free base pages are above 'mark'. For high-order checks it
2479 * will return true of the order-0 watermark is reached and there is at least
2480 * one free page of a suitable size. Checking now avoids taking the zone lock
2481 * to check in the allocation paths if no pages are free.
2483 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2484 unsigned long mark, int classzone_idx, int alloc_flags,
2489 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2491 /* free_pages may go negative - that's OK */
2492 free_pages -= (1 << order) - 1;
2494 if (alloc_flags & ALLOC_HIGH)
2498 * If the caller does not have rights to ALLOC_HARDER then subtract
2499 * the high-atomic reserves. This will over-estimate the size of the
2500 * atomic reserve but it avoids a search.
2502 if (likely(!alloc_harder))
2503 free_pages -= z->nr_reserved_highatomic;
2508 /* If allocation can't use CMA areas don't use free CMA pages */
2509 if (!(alloc_flags & ALLOC_CMA))
2510 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2514 * Check watermarks for an order-0 allocation request. If these
2515 * are not met, then a high-order request also cannot go ahead
2516 * even if a suitable page happened to be free.
2518 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2521 /* If this is an order-0 request then the watermark is fine */
2525 /* For a high-order request, check at least one suitable page is free */
2526 for (o = order; o < MAX_ORDER; o++) {
2527 struct free_area *area = &z->free_area[o];
2536 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2537 if (!list_empty(&area->free_list[mt]))
2542 if ((alloc_flags & ALLOC_CMA) &&
2543 !list_empty(&area->free_list[MIGRATE_CMA])) {
2551 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2552 int classzone_idx, int alloc_flags)
2554 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2555 zone_page_state(z, NR_FREE_PAGES));
2558 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2559 unsigned long mark, int classzone_idx)
2561 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2563 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2564 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2566 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2571 static bool zone_local(struct zone *local_zone, struct zone *zone)
2573 return local_zone->node == zone->node;
2576 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2578 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2581 #else /* CONFIG_NUMA */
2582 static bool zone_local(struct zone *local_zone, struct zone *zone)
2587 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2591 #endif /* CONFIG_NUMA */
2593 static void reset_alloc_batches(struct zone *preferred_zone)
2595 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2598 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2599 high_wmark_pages(zone) - low_wmark_pages(zone) -
2600 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2601 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2602 } while (zone++ != preferred_zone);
2606 * get_page_from_freelist goes through the zonelist trying to allocate
2609 static struct page *
2610 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2611 const struct alloc_context *ac)
2613 struct zonelist *zonelist = ac->zonelist;
2615 struct page *page = NULL;
2617 int nr_fair_skipped = 0;
2618 bool zonelist_rescan;
2621 zonelist_rescan = false;
2624 * Scan zonelist, looking for a zone with enough free.
2625 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2627 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2631 if (cpusets_enabled() &&
2632 (alloc_flags & ALLOC_CPUSET) &&
2633 !cpuset_zone_allowed(zone, gfp_mask))
2636 * Distribute pages in proportion to the individual
2637 * zone size to ensure fair page aging. The zone a
2638 * page was allocated in should have no effect on the
2639 * time the page has in memory before being reclaimed.
2641 if (alloc_flags & ALLOC_FAIR) {
2642 if (!zone_local(ac->preferred_zone, zone))
2644 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2650 * When allocating a page cache page for writing, we
2651 * want to get it from a zone that is within its dirty
2652 * limit, such that no single zone holds more than its
2653 * proportional share of globally allowed dirty pages.
2654 * The dirty limits take into account the zone's
2655 * lowmem reserves and high watermark so that kswapd
2656 * should be able to balance it without having to
2657 * write pages from its LRU list.
2659 * This may look like it could increase pressure on
2660 * lower zones by failing allocations in higher zones
2661 * before they are full. But the pages that do spill
2662 * over are limited as the lower zones are protected
2663 * by this very same mechanism. It should not become
2664 * a practical burden to them.
2666 * XXX: For now, allow allocations to potentially
2667 * exceed the per-zone dirty limit in the slowpath
2668 * (spread_dirty_pages unset) before going into reclaim,
2669 * which is important when on a NUMA setup the allowed
2670 * zones are together not big enough to reach the
2671 * global limit. The proper fix for these situations
2672 * will require awareness of zones in the
2673 * dirty-throttling and the flusher threads.
2675 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2678 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2679 if (!zone_watermark_ok(zone, order, mark,
2680 ac->classzone_idx, alloc_flags)) {
2683 /* Checked here to keep the fast path fast */
2684 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2685 if (alloc_flags & ALLOC_NO_WATERMARKS)
2688 if (zone_reclaim_mode == 0 ||
2689 !zone_allows_reclaim(ac->preferred_zone, zone))
2692 ret = zone_reclaim(zone, gfp_mask, order);
2694 case ZONE_RECLAIM_NOSCAN:
2697 case ZONE_RECLAIM_FULL:
2698 /* scanned but unreclaimable */
2701 /* did we reclaim enough */
2702 if (zone_watermark_ok(zone, order, mark,
2703 ac->classzone_idx, alloc_flags))
2711 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2712 gfp_mask, alloc_flags, ac->migratetype);
2714 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2718 * If this is a high-order atomic allocation then check
2719 * if the pageblock should be reserved for the future
2721 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2722 reserve_highatomic_pageblock(page, zone, order);
2729 * The first pass makes sure allocations are spread fairly within the
2730 * local node. However, the local node might have free pages left
2731 * after the fairness batches are exhausted, and remote zones haven't
2732 * even been considered yet. Try once more without fairness, and
2733 * include remote zones now, before entering the slowpath and waking
2734 * kswapd: prefer spilling to a remote zone over swapping locally.
2736 if (alloc_flags & ALLOC_FAIR) {
2737 alloc_flags &= ~ALLOC_FAIR;
2738 if (nr_fair_skipped) {
2739 zonelist_rescan = true;
2740 reset_alloc_batches(ac->preferred_zone);
2742 if (nr_online_nodes > 1)
2743 zonelist_rescan = true;
2746 if (zonelist_rescan)
2753 * Large machines with many possible nodes should not always dump per-node
2754 * meminfo in irq context.
2756 static inline bool should_suppress_show_mem(void)
2761 ret = in_interrupt();
2766 static DEFINE_RATELIMIT_STATE(nopage_rs,
2767 DEFAULT_RATELIMIT_INTERVAL,
2768 DEFAULT_RATELIMIT_BURST);
2770 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2772 unsigned int filter = SHOW_MEM_FILTER_NODES;
2774 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2775 debug_guardpage_minorder() > 0)
2779 * This documents exceptions given to allocations in certain
2780 * contexts that are allowed to allocate outside current's set
2783 if (!(gfp_mask & __GFP_NOMEMALLOC))
2784 if (test_thread_flag(TIF_MEMDIE) ||
2785 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2786 filter &= ~SHOW_MEM_FILTER_NODES;
2787 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2788 filter &= ~SHOW_MEM_FILTER_NODES;
2791 struct va_format vaf;
2794 va_start(args, fmt);
2799 pr_warn("%pV", &vaf);
2804 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2805 current->comm, order, gfp_mask, &gfp_mask);
2807 if (!should_suppress_show_mem())
2811 static inline struct page *
2812 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2813 const struct alloc_context *ac, unsigned long *did_some_progress)
2815 struct oom_control oc = {
2816 .zonelist = ac->zonelist,
2817 .nodemask = ac->nodemask,
2818 .gfp_mask = gfp_mask,
2823 *did_some_progress = 0;
2826 * Acquire the oom lock. If that fails, somebody else is
2827 * making progress for us.
2829 if (!mutex_trylock(&oom_lock)) {
2830 *did_some_progress = 1;
2831 schedule_timeout_uninterruptible(1);
2836 * Go through the zonelist yet one more time, keep very high watermark
2837 * here, this is only to catch a parallel oom killing, we must fail if
2838 * we're still under heavy pressure.
2840 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2841 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2845 if (!(gfp_mask & __GFP_NOFAIL)) {
2846 /* Coredumps can quickly deplete all memory reserves */
2847 if (current->flags & PF_DUMPCORE)
2849 /* The OOM killer will not help higher order allocs */
2850 if (order > PAGE_ALLOC_COSTLY_ORDER)
2852 /* The OOM killer does not needlessly kill tasks for lowmem */
2853 if (ac->high_zoneidx < ZONE_NORMAL)
2855 /* The OOM killer does not compensate for IO-less reclaim */
2856 if (!(gfp_mask & __GFP_FS)) {
2858 * XXX: Page reclaim didn't yield anything,
2859 * and the OOM killer can't be invoked, but
2860 * keep looping as per tradition.
2862 * But do not keep looping if oom_killer_disable()
2863 * was already called, for the system is trying to
2864 * enter a quiescent state during suspend.
2866 *did_some_progress = !oom_killer_disabled;
2869 if (pm_suspended_storage())
2871 /* The OOM killer may not free memory on a specific node */
2872 if (gfp_mask & __GFP_THISNODE)
2875 /* Exhausted what can be done so it's blamo time */
2876 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2877 *did_some_progress = 1;
2879 if (gfp_mask & __GFP_NOFAIL) {
2880 page = get_page_from_freelist(gfp_mask, order,
2881 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2883 * fallback to ignore cpuset restriction if our nodes
2887 page = get_page_from_freelist(gfp_mask, order,
2888 ALLOC_NO_WATERMARKS, ac);
2892 mutex_unlock(&oom_lock);
2896 #ifdef CONFIG_COMPACTION
2897 /* Try memory compaction for high-order allocations before reclaim */
2898 static struct page *
2899 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2900 int alloc_flags, const struct alloc_context *ac,
2901 enum migrate_mode mode, int *contended_compaction,
2902 bool *deferred_compaction)
2904 unsigned long compact_result;
2910 current->flags |= PF_MEMALLOC;
2911 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2912 mode, contended_compaction);
2913 current->flags &= ~PF_MEMALLOC;
2915 switch (compact_result) {
2916 case COMPACT_DEFERRED:
2917 *deferred_compaction = true;
2919 case COMPACT_SKIPPED:
2926 * At least in one zone compaction wasn't deferred or skipped, so let's
2927 * count a compaction stall
2929 count_vm_event(COMPACTSTALL);
2931 page = get_page_from_freelist(gfp_mask, order,
2932 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2935 struct zone *zone = page_zone(page);
2937 zone->compact_blockskip_flush = false;
2938 compaction_defer_reset(zone, order, true);
2939 count_vm_event(COMPACTSUCCESS);
2944 * It's bad if compaction run occurs and fails. The most likely reason
2945 * is that pages exist, but not enough to satisfy watermarks.
2947 count_vm_event(COMPACTFAIL);
2954 static inline struct page *
2955 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2956 int alloc_flags, const struct alloc_context *ac,
2957 enum migrate_mode mode, int *contended_compaction,
2958 bool *deferred_compaction)
2962 #endif /* CONFIG_COMPACTION */
2964 /* Perform direct synchronous page reclaim */
2966 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2967 const struct alloc_context *ac)
2969 struct reclaim_state reclaim_state;
2974 /* We now go into synchronous reclaim */
2975 cpuset_memory_pressure_bump();
2976 current->flags |= PF_MEMALLOC;
2977 lockdep_set_current_reclaim_state(gfp_mask);
2978 reclaim_state.reclaimed_slab = 0;
2979 current->reclaim_state = &reclaim_state;
2981 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2984 current->reclaim_state = NULL;
2985 lockdep_clear_current_reclaim_state();
2986 current->flags &= ~PF_MEMALLOC;
2993 /* The really slow allocator path where we enter direct reclaim */
2994 static inline struct page *
2995 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2996 int alloc_flags, const struct alloc_context *ac,
2997 unsigned long *did_some_progress)
2999 struct page *page = NULL;
3000 bool drained = false;
3002 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3003 if (unlikely(!(*did_some_progress)))
3007 page = get_page_from_freelist(gfp_mask, order,
3008 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3011 * If an allocation failed after direct reclaim, it could be because
3012 * pages are pinned on the per-cpu lists or in high alloc reserves.
3013 * Shrink them them and try again
3015 if (!page && !drained) {
3016 unreserve_highatomic_pageblock(ac);
3017 drain_all_pages(NULL);
3025 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3030 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3031 ac->high_zoneidx, ac->nodemask)
3032 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
3036 gfp_to_alloc_flags(gfp_t gfp_mask)
3038 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3040 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3041 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3044 * The caller may dip into page reserves a bit more if the caller
3045 * cannot run direct reclaim, or if the caller has realtime scheduling
3046 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3047 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3049 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3051 if (gfp_mask & __GFP_ATOMIC) {
3053 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3054 * if it can't schedule.
3056 if (!(gfp_mask & __GFP_NOMEMALLOC))
3057 alloc_flags |= ALLOC_HARDER;
3059 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3060 * comment for __cpuset_node_allowed().
3062 alloc_flags &= ~ALLOC_CPUSET;
3063 } else if (unlikely(rt_task(current)) && !in_interrupt())
3064 alloc_flags |= ALLOC_HARDER;
3066 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3067 if (gfp_mask & __GFP_MEMALLOC)
3068 alloc_flags |= ALLOC_NO_WATERMARKS;
3069 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3070 alloc_flags |= ALLOC_NO_WATERMARKS;
3071 else if (!in_interrupt() &&
3072 ((current->flags & PF_MEMALLOC) ||
3073 unlikely(test_thread_flag(TIF_MEMDIE))))
3074 alloc_flags |= ALLOC_NO_WATERMARKS;
3077 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3078 alloc_flags |= ALLOC_CMA;
3083 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3085 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3088 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3090 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3093 static inline struct page *
3094 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3095 struct alloc_context *ac)
3097 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3098 struct page *page = NULL;
3100 unsigned long pages_reclaimed = 0;
3101 unsigned long did_some_progress;
3102 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3103 bool deferred_compaction = false;
3104 int contended_compaction = COMPACT_CONTENDED_NONE;
3107 * In the slowpath, we sanity check order to avoid ever trying to
3108 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3109 * be using allocators in order of preference for an area that is
3112 if (order >= MAX_ORDER) {
3113 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3118 * We also sanity check to catch abuse of atomic reserves being used by
3119 * callers that are not in atomic context.
3121 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3122 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3123 gfp_mask &= ~__GFP_ATOMIC;
3126 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3127 wake_all_kswapds(order, ac);
3130 * OK, we're below the kswapd watermark and have kicked background
3131 * reclaim. Now things get more complex, so set up alloc_flags according
3132 * to how we want to proceed.
3134 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3137 * Find the true preferred zone if the allocation is unconstrained by
3140 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3141 struct zoneref *preferred_zoneref;
3142 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3143 ac->high_zoneidx, NULL, &ac->preferred_zone);
3144 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3147 /* This is the last chance, in general, before the goto nopage. */
3148 page = get_page_from_freelist(gfp_mask, order,
3149 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3153 /* Allocate without watermarks if the context allows */
3154 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3156 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3157 * the allocation is high priority and these type of
3158 * allocations are system rather than user orientated
3160 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3161 page = get_page_from_freelist(gfp_mask, order,
3162 ALLOC_NO_WATERMARKS, ac);
3167 /* Caller is not willing to reclaim, we can't balance anything */
3168 if (!can_direct_reclaim) {
3170 * All existing users of the __GFP_NOFAIL are blockable, so warn
3171 * of any new users that actually allow this type of allocation
3174 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3178 /* Avoid recursion of direct reclaim */
3179 if (current->flags & PF_MEMALLOC) {
3181 * __GFP_NOFAIL request from this context is rather bizarre
3182 * because we cannot reclaim anything and only can loop waiting
3183 * for somebody to do a work for us.
3185 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3192 /* Avoid allocations with no watermarks from looping endlessly */
3193 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3197 * Try direct compaction. The first pass is asynchronous. Subsequent
3198 * attempts after direct reclaim are synchronous
3200 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3202 &contended_compaction,
3203 &deferred_compaction);
3207 /* Checks for THP-specific high-order allocations */
3208 if (is_thp_gfp_mask(gfp_mask)) {
3210 * If compaction is deferred for high-order allocations, it is
3211 * because sync compaction recently failed. If this is the case
3212 * and the caller requested a THP allocation, we do not want
3213 * to heavily disrupt the system, so we fail the allocation
3214 * instead of entering direct reclaim.
3216 if (deferred_compaction)
3220 * In all zones where compaction was attempted (and not
3221 * deferred or skipped), lock contention has been detected.
3222 * For THP allocation we do not want to disrupt the others
3223 * so we fallback to base pages instead.
3225 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3229 * If compaction was aborted due to need_resched(), we do not
3230 * want to further increase allocation latency, unless it is
3231 * khugepaged trying to collapse.
3233 if (contended_compaction == COMPACT_CONTENDED_SCHED
3234 && !(current->flags & PF_KTHREAD))
3239 * It can become very expensive to allocate transparent hugepages at
3240 * fault, so use asynchronous memory compaction for THP unless it is
3241 * khugepaged trying to collapse.
3243 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3244 migration_mode = MIGRATE_SYNC_LIGHT;
3246 /* Try direct reclaim and then allocating */
3247 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3248 &did_some_progress);
3252 /* Do not loop if specifically requested */
3253 if (gfp_mask & __GFP_NORETRY)
3256 /* Keep reclaiming pages as long as there is reasonable progress */
3257 pages_reclaimed += did_some_progress;
3258 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3259 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3260 /* Wait for some write requests to complete then retry */
3261 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3265 /* Reclaim has failed us, start killing things */
3266 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3270 /* Retry as long as the OOM killer is making progress */
3271 if (did_some_progress)
3276 * High-order allocations do not necessarily loop after
3277 * direct reclaim and reclaim/compaction depends on compaction
3278 * being called after reclaim so call directly if necessary
3280 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3282 &contended_compaction,
3283 &deferred_compaction);
3287 warn_alloc_failed(gfp_mask, order, NULL);
3293 * This is the 'heart' of the zoned buddy allocator.
3296 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3297 struct zonelist *zonelist, nodemask_t *nodemask)
3299 struct zoneref *preferred_zoneref;
3300 struct page *page = NULL;
3301 unsigned int cpuset_mems_cookie;
3302 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3303 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3304 struct alloc_context ac = {
3305 .high_zoneidx = gfp_zone(gfp_mask),
3306 .nodemask = nodemask,
3307 .migratetype = gfpflags_to_migratetype(gfp_mask),
3310 gfp_mask &= gfp_allowed_mask;
3312 lockdep_trace_alloc(gfp_mask);
3314 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3316 if (should_fail_alloc_page(gfp_mask, order))
3320 * Check the zones suitable for the gfp_mask contain at least one
3321 * valid zone. It's possible to have an empty zonelist as a result
3322 * of __GFP_THISNODE and a memoryless node
3324 if (unlikely(!zonelist->_zonerefs->zone))
3327 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3328 alloc_flags |= ALLOC_CMA;
3331 cpuset_mems_cookie = read_mems_allowed_begin();
3333 /* We set it here, as __alloc_pages_slowpath might have changed it */
3334 ac.zonelist = zonelist;
3336 /* Dirty zone balancing only done in the fast path */
3337 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3339 /* The preferred zone is used for statistics later */
3340 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3341 ac.nodemask ? : &cpuset_current_mems_allowed,
3342 &ac.preferred_zone);
3343 if (!ac.preferred_zone)
3345 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3347 /* First allocation attempt */
3348 alloc_mask = gfp_mask|__GFP_HARDWALL;
3349 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3350 if (unlikely(!page)) {
3352 * Runtime PM, block IO and its error handling path
3353 * can deadlock because I/O on the device might not
3356 alloc_mask = memalloc_noio_flags(gfp_mask);
3357 ac.spread_dirty_pages = false;
3359 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3362 if (kmemcheck_enabled && page)
3363 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3365 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3369 * When updating a task's mems_allowed, it is possible to race with
3370 * parallel threads in such a way that an allocation can fail while
3371 * the mask is being updated. If a page allocation is about to fail,
3372 * check if the cpuset changed during allocation and if so, retry.
3374 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3379 EXPORT_SYMBOL(__alloc_pages_nodemask);
3382 * Common helper functions.
3384 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3389 * __get_free_pages() returns a 32-bit address, which cannot represent
3392 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3394 page = alloc_pages(gfp_mask, order);
3397 return (unsigned long) page_address(page);
3399 EXPORT_SYMBOL(__get_free_pages);
3401 unsigned long get_zeroed_page(gfp_t gfp_mask)
3403 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3405 EXPORT_SYMBOL(get_zeroed_page);
3407 void __free_pages(struct page *page, unsigned int order)
3409 if (put_page_testzero(page)) {
3411 free_hot_cold_page(page, false);
3413 __free_pages_ok(page, order);
3417 EXPORT_SYMBOL(__free_pages);
3419 void free_pages(unsigned long addr, unsigned int order)
3422 VM_BUG_ON(!virt_addr_valid((void *)addr));
3423 __free_pages(virt_to_page((void *)addr), order);
3427 EXPORT_SYMBOL(free_pages);
3431 * An arbitrary-length arbitrary-offset area of memory which resides
3432 * within a 0 or higher order page. Multiple fragments within that page
3433 * are individually refcounted, in the page's reference counter.
3435 * The page_frag functions below provide a simple allocation framework for
3436 * page fragments. This is used by the network stack and network device
3437 * drivers to provide a backing region of memory for use as either an
3438 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3440 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3443 struct page *page = NULL;
3444 gfp_t gfp = gfp_mask;
3446 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3447 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3449 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3450 PAGE_FRAG_CACHE_MAX_ORDER);
3451 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3453 if (unlikely(!page))
3454 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3456 nc->va = page ? page_address(page) : NULL;
3461 void *__alloc_page_frag(struct page_frag_cache *nc,
3462 unsigned int fragsz, gfp_t gfp_mask)
3464 unsigned int size = PAGE_SIZE;
3468 if (unlikely(!nc->va)) {
3470 page = __page_frag_refill(nc, gfp_mask);
3474 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3475 /* if size can vary use size else just use PAGE_SIZE */
3478 /* Even if we own the page, we do not use atomic_set().
3479 * This would break get_page_unless_zero() users.
3481 page_ref_add(page, size - 1);
3483 /* reset page count bias and offset to start of new frag */
3484 nc->pfmemalloc = page_is_pfmemalloc(page);
3485 nc->pagecnt_bias = size;
3489 offset = nc->offset - fragsz;
3490 if (unlikely(offset < 0)) {
3491 page = virt_to_page(nc->va);
3493 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3496 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3497 /* if size can vary use size else just use PAGE_SIZE */
3500 /* OK, page count is 0, we can safely set it */
3501 set_page_count(page, size);
3503 /* reset page count bias and offset to start of new frag */
3504 nc->pagecnt_bias = size;
3505 offset = size - fragsz;
3509 nc->offset = offset;
3511 return nc->va + offset;
3513 EXPORT_SYMBOL(__alloc_page_frag);
3516 * Frees a page fragment allocated out of either a compound or order 0 page.
3518 void __free_page_frag(void *addr)
3520 struct page *page = virt_to_head_page(addr);
3522 if (unlikely(put_page_testzero(page)))
3523 __free_pages_ok(page, compound_order(page));
3525 EXPORT_SYMBOL(__free_page_frag);
3528 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3529 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3530 * equivalent to alloc_pages.
3532 * It should be used when the caller would like to use kmalloc, but since the
3533 * allocation is large, it has to fall back to the page allocator.
3535 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3539 page = alloc_pages(gfp_mask, order);
3540 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3541 __free_pages(page, order);
3547 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3551 page = alloc_pages_node(nid, gfp_mask, order);
3552 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3553 __free_pages(page, order);
3560 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3563 void __free_kmem_pages(struct page *page, unsigned int order)
3565 memcg_kmem_uncharge(page, order);
3566 __free_pages(page, order);
3569 void free_kmem_pages(unsigned long addr, unsigned int order)
3572 VM_BUG_ON(!virt_addr_valid((void *)addr));
3573 __free_kmem_pages(virt_to_page((void *)addr), order);
3577 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3581 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3582 unsigned long used = addr + PAGE_ALIGN(size);
3584 split_page(virt_to_page((void *)addr), order);
3585 while (used < alloc_end) {
3590 return (void *)addr;
3594 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3595 * @size: the number of bytes to allocate
3596 * @gfp_mask: GFP flags for the allocation
3598 * This function is similar to alloc_pages(), except that it allocates the
3599 * minimum number of pages to satisfy the request. alloc_pages() can only
3600 * allocate memory in power-of-two pages.
3602 * This function is also limited by MAX_ORDER.
3604 * Memory allocated by this function must be released by free_pages_exact().
3606 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3608 unsigned int order = get_order(size);
3611 addr = __get_free_pages(gfp_mask, order);
3612 return make_alloc_exact(addr, order, size);
3614 EXPORT_SYMBOL(alloc_pages_exact);
3617 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3619 * @nid: the preferred node ID where memory should be allocated
3620 * @size: the number of bytes to allocate
3621 * @gfp_mask: GFP flags for the allocation
3623 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3626 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3628 unsigned int order = get_order(size);
3629 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3632 return make_alloc_exact((unsigned long)page_address(p), order, size);
3636 * free_pages_exact - release memory allocated via alloc_pages_exact()
3637 * @virt: the value returned by alloc_pages_exact.
3638 * @size: size of allocation, same value as passed to alloc_pages_exact().
3640 * Release the memory allocated by a previous call to alloc_pages_exact.
3642 void free_pages_exact(void *virt, size_t size)
3644 unsigned long addr = (unsigned long)virt;
3645 unsigned long end = addr + PAGE_ALIGN(size);
3647 while (addr < end) {
3652 EXPORT_SYMBOL(free_pages_exact);
3655 * nr_free_zone_pages - count number of pages beyond high watermark
3656 * @offset: The zone index of the highest zone
3658 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3659 * high watermark within all zones at or below a given zone index. For each
3660 * zone, the number of pages is calculated as:
3661 * managed_pages - high_pages
3663 static unsigned long nr_free_zone_pages(int offset)
3668 /* Just pick one node, since fallback list is circular */
3669 unsigned long sum = 0;
3671 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3673 for_each_zone_zonelist(zone, z, zonelist, offset) {
3674 unsigned long size = zone->managed_pages;
3675 unsigned long high = high_wmark_pages(zone);
3684 * nr_free_buffer_pages - count number of pages beyond high watermark
3686 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3687 * watermark within ZONE_DMA and ZONE_NORMAL.
3689 unsigned long nr_free_buffer_pages(void)
3691 return nr_free_zone_pages(gfp_zone(GFP_USER));
3693 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3696 * nr_free_pagecache_pages - count number of pages beyond high watermark
3698 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3699 * high watermark within all zones.
3701 unsigned long nr_free_pagecache_pages(void)
3703 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3706 static inline void show_node(struct zone *zone)
3708 if (IS_ENABLED(CONFIG_NUMA))
3709 printk("Node %d ", zone_to_nid(zone));
3712 long si_mem_available(void)
3715 unsigned long pagecache;
3716 unsigned long wmark_low = 0;
3717 unsigned long pages[NR_LRU_LISTS];
3721 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3722 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3725 wmark_low += zone->watermark[WMARK_LOW];
3728 * Estimate the amount of memory available for userspace allocations,
3729 * without causing swapping.
3731 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3734 * Not all the page cache can be freed, otherwise the system will
3735 * start swapping. Assume at least half of the page cache, or the
3736 * low watermark worth of cache, needs to stay.
3738 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3739 pagecache -= min(pagecache / 2, wmark_low);
3740 available += pagecache;
3743 * Part of the reclaimable slab consists of items that are in use,
3744 * and cannot be freed. Cap this estimate at the low watermark.
3746 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3747 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3753 EXPORT_SYMBOL_GPL(si_mem_available);
3755 void si_meminfo(struct sysinfo *val)
3757 val->totalram = totalram_pages;
3758 val->sharedram = global_page_state(NR_SHMEM);
3759 val->freeram = global_page_state(NR_FREE_PAGES);
3760 val->bufferram = nr_blockdev_pages();
3761 val->totalhigh = totalhigh_pages;
3762 val->freehigh = nr_free_highpages();
3763 val->mem_unit = PAGE_SIZE;
3766 EXPORT_SYMBOL(si_meminfo);
3769 void si_meminfo_node(struct sysinfo *val, int nid)
3771 int zone_type; /* needs to be signed */
3772 unsigned long managed_pages = 0;
3773 pg_data_t *pgdat = NODE_DATA(nid);
3775 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3776 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3777 val->totalram = managed_pages;
3778 val->sharedram = node_page_state(nid, NR_SHMEM);
3779 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3780 #ifdef CONFIG_HIGHMEM
3781 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3782 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3788 val->mem_unit = PAGE_SIZE;
3793 * Determine whether the node should be displayed or not, depending on whether
3794 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3796 bool skip_free_areas_node(unsigned int flags, int nid)
3799 unsigned int cpuset_mems_cookie;
3801 if (!(flags & SHOW_MEM_FILTER_NODES))
3805 cpuset_mems_cookie = read_mems_allowed_begin();
3806 ret = !node_isset(nid, cpuset_current_mems_allowed);
3807 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3812 #define K(x) ((x) << (PAGE_SHIFT-10))
3814 static void show_migration_types(unsigned char type)
3816 static const char types[MIGRATE_TYPES] = {
3817 [MIGRATE_UNMOVABLE] = 'U',
3818 [MIGRATE_MOVABLE] = 'M',
3819 [MIGRATE_RECLAIMABLE] = 'E',
3820 [MIGRATE_HIGHATOMIC] = 'H',
3822 [MIGRATE_CMA] = 'C',
3824 #ifdef CONFIG_MEMORY_ISOLATION
3825 [MIGRATE_ISOLATE] = 'I',
3828 char tmp[MIGRATE_TYPES + 1];
3832 for (i = 0; i < MIGRATE_TYPES; i++) {
3833 if (type & (1 << i))
3838 printk("(%s) ", tmp);
3842 * Show free area list (used inside shift_scroll-lock stuff)
3843 * We also calculate the percentage fragmentation. We do this by counting the
3844 * memory on each free list with the exception of the first item on the list.
3847 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3850 void show_free_areas(unsigned int filter)
3852 unsigned long free_pcp = 0;
3856 for_each_populated_zone(zone) {
3857 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3860 for_each_online_cpu(cpu)
3861 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3864 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3865 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3866 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3867 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3868 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3869 " free:%lu free_pcp:%lu free_cma:%lu\n",
3870 global_page_state(NR_ACTIVE_ANON),
3871 global_page_state(NR_INACTIVE_ANON),
3872 global_page_state(NR_ISOLATED_ANON),
3873 global_page_state(NR_ACTIVE_FILE),
3874 global_page_state(NR_INACTIVE_FILE),
3875 global_page_state(NR_ISOLATED_FILE),
3876 global_page_state(NR_UNEVICTABLE),
3877 global_page_state(NR_FILE_DIRTY),
3878 global_page_state(NR_WRITEBACK),
3879 global_page_state(NR_UNSTABLE_NFS),
3880 global_page_state(NR_SLAB_RECLAIMABLE),
3881 global_page_state(NR_SLAB_UNRECLAIMABLE),
3882 global_page_state(NR_FILE_MAPPED),
3883 global_page_state(NR_SHMEM),
3884 global_page_state(NR_PAGETABLE),
3885 global_page_state(NR_BOUNCE),
3886 global_page_state(NR_FREE_PAGES),
3888 global_page_state(NR_FREE_CMA_PAGES));
3890 for_each_populated_zone(zone) {
3893 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3897 for_each_online_cpu(cpu)
3898 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3906 " active_anon:%lukB"
3907 " inactive_anon:%lukB"
3908 " active_file:%lukB"
3909 " inactive_file:%lukB"
3910 " unevictable:%lukB"
3911 " isolated(anon):%lukB"
3912 " isolated(file):%lukB"
3920 " slab_reclaimable:%lukB"
3921 " slab_unreclaimable:%lukB"
3922 " kernel_stack:%lukB"
3929 " writeback_tmp:%lukB"
3930 " pages_scanned:%lu"
3931 " all_unreclaimable? %s"
3934 K(zone_page_state(zone, NR_FREE_PAGES)),
3935 K(min_wmark_pages(zone)),
3936 K(low_wmark_pages(zone)),
3937 K(high_wmark_pages(zone)),
3938 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3939 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3940 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3941 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3942 K(zone_page_state(zone, NR_UNEVICTABLE)),
3943 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3944 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3945 K(zone->present_pages),
3946 K(zone->managed_pages),
3947 K(zone_page_state(zone, NR_MLOCK)),
3948 K(zone_page_state(zone, NR_FILE_DIRTY)),
3949 K(zone_page_state(zone, NR_WRITEBACK)),
3950 K(zone_page_state(zone, NR_FILE_MAPPED)),
3951 K(zone_page_state(zone, NR_SHMEM)),
3952 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3953 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3954 zone_page_state(zone, NR_KERNEL_STACK) *
3956 K(zone_page_state(zone, NR_PAGETABLE)),
3957 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3958 K(zone_page_state(zone, NR_BOUNCE)),
3960 K(this_cpu_read(zone->pageset->pcp.count)),
3961 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3962 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3963 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3964 (!zone_reclaimable(zone) ? "yes" : "no")
3966 printk("lowmem_reserve[]:");
3967 for (i = 0; i < MAX_NR_ZONES; i++)
3968 printk(" %ld", zone->lowmem_reserve[i]);
3972 for_each_populated_zone(zone) {
3974 unsigned long nr[MAX_ORDER], flags, total = 0;
3975 unsigned char types[MAX_ORDER];
3977 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3980 printk("%s: ", zone->name);
3982 spin_lock_irqsave(&zone->lock, flags);
3983 for (order = 0; order < MAX_ORDER; order++) {
3984 struct free_area *area = &zone->free_area[order];
3987 nr[order] = area->nr_free;
3988 total += nr[order] << order;
3991 for (type = 0; type < MIGRATE_TYPES; type++) {
3992 if (!list_empty(&area->free_list[type]))
3993 types[order] |= 1 << type;
3996 spin_unlock_irqrestore(&zone->lock, flags);
3997 for (order = 0; order < MAX_ORDER; order++) {
3998 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4000 show_migration_types(types[order]);
4002 printk("= %lukB\n", K(total));
4005 hugetlb_show_meminfo();
4007 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4009 show_swap_cache_info();
4012 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4014 zoneref->zone = zone;
4015 zoneref->zone_idx = zone_idx(zone);
4019 * Builds allocation fallback zone lists.
4021 * Add all populated zones of a node to the zonelist.
4023 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4027 enum zone_type zone_type = MAX_NR_ZONES;
4031 zone = pgdat->node_zones + zone_type;
4032 if (populated_zone(zone)) {
4033 zoneref_set_zone(zone,
4034 &zonelist->_zonerefs[nr_zones++]);
4035 check_highest_zone(zone_type);
4037 } while (zone_type);
4045 * 0 = automatic detection of better ordering.
4046 * 1 = order by ([node] distance, -zonetype)
4047 * 2 = order by (-zonetype, [node] distance)
4049 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4050 * the same zonelist. So only NUMA can configure this param.
4052 #define ZONELIST_ORDER_DEFAULT 0
4053 #define ZONELIST_ORDER_NODE 1
4054 #define ZONELIST_ORDER_ZONE 2
4056 /* zonelist order in the kernel.
4057 * set_zonelist_order() will set this to NODE or ZONE.
4059 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4060 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4064 /* The value user specified ....changed by config */
4065 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4066 /* string for sysctl */
4067 #define NUMA_ZONELIST_ORDER_LEN 16
4068 char numa_zonelist_order[16] = "default";
4071 * interface for configure zonelist ordering.
4072 * command line option "numa_zonelist_order"
4073 * = "[dD]efault - default, automatic configuration.
4074 * = "[nN]ode - order by node locality, then by zone within node
4075 * = "[zZ]one - order by zone, then by locality within zone
4078 static int __parse_numa_zonelist_order(char *s)
4080 if (*s == 'd' || *s == 'D') {
4081 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4082 } else if (*s == 'n' || *s == 'N') {
4083 user_zonelist_order = ZONELIST_ORDER_NODE;
4084 } else if (*s == 'z' || *s == 'Z') {
4085 user_zonelist_order = ZONELIST_ORDER_ZONE;
4087 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4093 static __init int setup_numa_zonelist_order(char *s)
4100 ret = __parse_numa_zonelist_order(s);
4102 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4106 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4109 * sysctl handler for numa_zonelist_order
4111 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4112 void __user *buffer, size_t *length,
4115 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4117 static DEFINE_MUTEX(zl_order_mutex);
4119 mutex_lock(&zl_order_mutex);
4121 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4125 strcpy(saved_string, (char *)table->data);
4127 ret = proc_dostring(table, write, buffer, length, ppos);
4131 int oldval = user_zonelist_order;
4133 ret = __parse_numa_zonelist_order((char *)table->data);
4136 * bogus value. restore saved string
4138 strncpy((char *)table->data, saved_string,
4139 NUMA_ZONELIST_ORDER_LEN);
4140 user_zonelist_order = oldval;
4141 } else if (oldval != user_zonelist_order) {
4142 mutex_lock(&zonelists_mutex);
4143 build_all_zonelists(NULL, NULL);
4144 mutex_unlock(&zonelists_mutex);
4148 mutex_unlock(&zl_order_mutex);
4153 #define MAX_NODE_LOAD (nr_online_nodes)
4154 static int node_load[MAX_NUMNODES];
4157 * find_next_best_node - find the next node that should appear in a given node's fallback list
4158 * @node: node whose fallback list we're appending
4159 * @used_node_mask: nodemask_t of already used nodes
4161 * We use a number of factors to determine which is the next node that should
4162 * appear on a given node's fallback list. The node should not have appeared
4163 * already in @node's fallback list, and it should be the next closest node
4164 * according to the distance array (which contains arbitrary distance values
4165 * from each node to each node in the system), and should also prefer nodes
4166 * with no CPUs, since presumably they'll have very little allocation pressure
4167 * on them otherwise.
4168 * It returns -1 if no node is found.
4170 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4173 int min_val = INT_MAX;
4174 int best_node = NUMA_NO_NODE;
4175 const struct cpumask *tmp = cpumask_of_node(0);
4177 /* Use the local node if we haven't already */
4178 if (!node_isset(node, *used_node_mask)) {
4179 node_set(node, *used_node_mask);
4183 for_each_node_state(n, N_MEMORY) {
4185 /* Don't want a node to appear more than once */
4186 if (node_isset(n, *used_node_mask))
4189 /* Use the distance array to find the distance */
4190 val = node_distance(node, n);
4192 /* Penalize nodes under us ("prefer the next node") */
4195 /* Give preference to headless and unused nodes */
4196 tmp = cpumask_of_node(n);
4197 if (!cpumask_empty(tmp))
4198 val += PENALTY_FOR_NODE_WITH_CPUS;
4200 /* Slight preference for less loaded node */
4201 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4202 val += node_load[n];
4204 if (val < min_val) {
4211 node_set(best_node, *used_node_mask);
4218 * Build zonelists ordered by node and zones within node.
4219 * This results in maximum locality--normal zone overflows into local
4220 * DMA zone, if any--but risks exhausting DMA zone.
4222 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4225 struct zonelist *zonelist;
4227 zonelist = &pgdat->node_zonelists[0];
4228 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4230 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4231 zonelist->_zonerefs[j].zone = NULL;
4232 zonelist->_zonerefs[j].zone_idx = 0;
4236 * Build gfp_thisnode zonelists
4238 static void build_thisnode_zonelists(pg_data_t *pgdat)
4241 struct zonelist *zonelist;
4243 zonelist = &pgdat->node_zonelists[1];
4244 j = build_zonelists_node(pgdat, zonelist, 0);
4245 zonelist->_zonerefs[j].zone = NULL;
4246 zonelist->_zonerefs[j].zone_idx = 0;
4250 * Build zonelists ordered by zone and nodes within zones.
4251 * This results in conserving DMA zone[s] until all Normal memory is
4252 * exhausted, but results in overflowing to remote node while memory
4253 * may still exist in local DMA zone.
4255 static int node_order[MAX_NUMNODES];
4257 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4260 int zone_type; /* needs to be signed */
4262 struct zonelist *zonelist;
4264 zonelist = &pgdat->node_zonelists[0];
4266 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4267 for (j = 0; j < nr_nodes; j++) {
4268 node = node_order[j];
4269 z = &NODE_DATA(node)->node_zones[zone_type];
4270 if (populated_zone(z)) {
4272 &zonelist->_zonerefs[pos++]);
4273 check_highest_zone(zone_type);
4277 zonelist->_zonerefs[pos].zone = NULL;
4278 zonelist->_zonerefs[pos].zone_idx = 0;
4281 #if defined(CONFIG_64BIT)
4283 * Devices that require DMA32/DMA are relatively rare and do not justify a
4284 * penalty to every machine in case the specialised case applies. Default
4285 * to Node-ordering on 64-bit NUMA machines
4287 static int default_zonelist_order(void)
4289 return ZONELIST_ORDER_NODE;
4293 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4294 * by the kernel. If processes running on node 0 deplete the low memory zone
4295 * then reclaim will occur more frequency increasing stalls and potentially
4296 * be easier to OOM if a large percentage of the zone is under writeback or
4297 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4298 * Hence, default to zone ordering on 32-bit.
4300 static int default_zonelist_order(void)
4302 return ZONELIST_ORDER_ZONE;
4304 #endif /* CONFIG_64BIT */
4306 static void set_zonelist_order(void)
4308 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4309 current_zonelist_order = default_zonelist_order();
4311 current_zonelist_order = user_zonelist_order;
4314 static void build_zonelists(pg_data_t *pgdat)
4317 nodemask_t used_mask;
4318 int local_node, prev_node;
4319 struct zonelist *zonelist;
4320 unsigned int order = current_zonelist_order;
4322 /* initialize zonelists */
4323 for (i = 0; i < MAX_ZONELISTS; i++) {
4324 zonelist = pgdat->node_zonelists + i;
4325 zonelist->_zonerefs[0].zone = NULL;
4326 zonelist->_zonerefs[0].zone_idx = 0;
4329 /* NUMA-aware ordering of nodes */
4330 local_node = pgdat->node_id;
4331 load = nr_online_nodes;
4332 prev_node = local_node;
4333 nodes_clear(used_mask);
4335 memset(node_order, 0, sizeof(node_order));
4338 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4340 * We don't want to pressure a particular node.
4341 * So adding penalty to the first node in same
4342 * distance group to make it round-robin.
4344 if (node_distance(local_node, node) !=
4345 node_distance(local_node, prev_node))
4346 node_load[node] = load;
4350 if (order == ZONELIST_ORDER_NODE)
4351 build_zonelists_in_node_order(pgdat, node);
4353 node_order[i++] = node; /* remember order */
4356 if (order == ZONELIST_ORDER_ZONE) {
4357 /* calculate node order -- i.e., DMA last! */
4358 build_zonelists_in_zone_order(pgdat, i);
4361 build_thisnode_zonelists(pgdat);
4364 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4366 * Return node id of node used for "local" allocations.
4367 * I.e., first node id of first zone in arg node's generic zonelist.
4368 * Used for initializing percpu 'numa_mem', which is used primarily
4369 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4371 int local_memory_node(int node)
4375 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4376 gfp_zone(GFP_KERNEL),
4383 #else /* CONFIG_NUMA */
4385 static void set_zonelist_order(void)
4387 current_zonelist_order = ZONELIST_ORDER_ZONE;
4390 static void build_zonelists(pg_data_t *pgdat)
4392 int node, local_node;
4394 struct zonelist *zonelist;
4396 local_node = pgdat->node_id;
4398 zonelist = &pgdat->node_zonelists[0];
4399 j = build_zonelists_node(pgdat, zonelist, 0);
4402 * Now we build the zonelist so that it contains the zones
4403 * of all the other nodes.
4404 * We don't want to pressure a particular node, so when
4405 * building the zones for node N, we make sure that the
4406 * zones coming right after the local ones are those from
4407 * node N+1 (modulo N)
4409 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4410 if (!node_online(node))
4412 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4414 for (node = 0; node < local_node; node++) {
4415 if (!node_online(node))
4417 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4420 zonelist->_zonerefs[j].zone = NULL;
4421 zonelist->_zonerefs[j].zone_idx = 0;
4424 #endif /* CONFIG_NUMA */
4427 * Boot pageset table. One per cpu which is going to be used for all
4428 * zones and all nodes. The parameters will be set in such a way
4429 * that an item put on a list will immediately be handed over to
4430 * the buddy list. This is safe since pageset manipulation is done
4431 * with interrupts disabled.
4433 * The boot_pagesets must be kept even after bootup is complete for
4434 * unused processors and/or zones. They do play a role for bootstrapping
4435 * hotplugged processors.
4437 * zoneinfo_show() and maybe other functions do
4438 * not check if the processor is online before following the pageset pointer.
4439 * Other parts of the kernel may not check if the zone is available.
4441 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4442 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4443 static void setup_zone_pageset(struct zone *zone);
4446 * Global mutex to protect against size modification of zonelists
4447 * as well as to serialize pageset setup for the new populated zone.
4449 DEFINE_MUTEX(zonelists_mutex);
4451 /* return values int ....just for stop_machine() */
4452 static int __build_all_zonelists(void *data)
4456 pg_data_t *self = data;
4459 memset(node_load, 0, sizeof(node_load));
4462 if (self && !node_online(self->node_id)) {
4463 build_zonelists(self);
4466 for_each_online_node(nid) {
4467 pg_data_t *pgdat = NODE_DATA(nid);
4469 build_zonelists(pgdat);
4473 * Initialize the boot_pagesets that are going to be used
4474 * for bootstrapping processors. The real pagesets for
4475 * each zone will be allocated later when the per cpu
4476 * allocator is available.
4478 * boot_pagesets are used also for bootstrapping offline
4479 * cpus if the system is already booted because the pagesets
4480 * are needed to initialize allocators on a specific cpu too.
4481 * F.e. the percpu allocator needs the page allocator which
4482 * needs the percpu allocator in order to allocate its pagesets
4483 * (a chicken-egg dilemma).
4485 for_each_possible_cpu(cpu) {
4486 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4488 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4490 * We now know the "local memory node" for each node--
4491 * i.e., the node of the first zone in the generic zonelist.
4492 * Set up numa_mem percpu variable for on-line cpus. During
4493 * boot, only the boot cpu should be on-line; we'll init the
4494 * secondary cpus' numa_mem as they come on-line. During
4495 * node/memory hotplug, we'll fixup all on-line cpus.
4497 if (cpu_online(cpu))
4498 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4505 static noinline void __init
4506 build_all_zonelists_init(void)
4508 __build_all_zonelists(NULL);
4509 mminit_verify_zonelist();
4510 cpuset_init_current_mems_allowed();
4514 * Called with zonelists_mutex held always
4515 * unless system_state == SYSTEM_BOOTING.
4517 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4518 * [we're only called with non-NULL zone through __meminit paths] and
4519 * (2) call of __init annotated helper build_all_zonelists_init
4520 * [protected by SYSTEM_BOOTING].
4522 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4524 set_zonelist_order();
4526 if (system_state == SYSTEM_BOOTING) {
4527 build_all_zonelists_init();
4529 #ifdef CONFIG_MEMORY_HOTPLUG
4531 setup_zone_pageset(zone);
4533 /* we have to stop all cpus to guarantee there is no user
4535 stop_machine(__build_all_zonelists, pgdat, NULL);
4536 /* cpuset refresh routine should be here */
4538 vm_total_pages = nr_free_pagecache_pages();
4540 * Disable grouping by mobility if the number of pages in the
4541 * system is too low to allow the mechanism to work. It would be
4542 * more accurate, but expensive to check per-zone. This check is
4543 * made on memory-hotadd so a system can start with mobility
4544 * disabled and enable it later
4546 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4547 page_group_by_mobility_disabled = 1;
4549 page_group_by_mobility_disabled = 0;
4551 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4553 zonelist_order_name[current_zonelist_order],
4554 page_group_by_mobility_disabled ? "off" : "on",
4557 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4562 * Helper functions to size the waitqueue hash table.
4563 * Essentially these want to choose hash table sizes sufficiently
4564 * large so that collisions trying to wait on pages are rare.
4565 * But in fact, the number of active page waitqueues on typical
4566 * systems is ridiculously low, less than 200. So this is even
4567 * conservative, even though it seems large.
4569 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4570 * waitqueues, i.e. the size of the waitq table given the number of pages.
4572 #define PAGES_PER_WAITQUEUE 256
4574 #ifndef CONFIG_MEMORY_HOTPLUG
4575 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4577 unsigned long size = 1;
4579 pages /= PAGES_PER_WAITQUEUE;
4581 while (size < pages)
4585 * Once we have dozens or even hundreds of threads sleeping
4586 * on IO we've got bigger problems than wait queue collision.
4587 * Limit the size of the wait table to a reasonable size.
4589 size = min(size, 4096UL);
4591 return max(size, 4UL);
4595 * A zone's size might be changed by hot-add, so it is not possible to determine
4596 * a suitable size for its wait_table. So we use the maximum size now.
4598 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4600 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4601 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4602 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4604 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4605 * or more by the traditional way. (See above). It equals:
4607 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4608 * ia64(16K page size) : = ( 8G + 4M)byte.
4609 * powerpc (64K page size) : = (32G +16M)byte.
4611 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4618 * This is an integer logarithm so that shifts can be used later
4619 * to extract the more random high bits from the multiplicative
4620 * hash function before the remainder is taken.
4622 static inline unsigned long wait_table_bits(unsigned long size)
4628 * Initially all pages are reserved - free ones are freed
4629 * up by free_all_bootmem() once the early boot process is
4630 * done. Non-atomic initialization, single-pass.
4632 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4633 unsigned long start_pfn, enum memmap_context context)
4635 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4636 unsigned long end_pfn = start_pfn + size;
4637 pg_data_t *pgdat = NODE_DATA(nid);
4639 unsigned long nr_initialised = 0;
4640 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4641 struct memblock_region *r = NULL, *tmp;
4644 if (highest_memmap_pfn < end_pfn - 1)
4645 highest_memmap_pfn = end_pfn - 1;
4648 * Honor reservation requested by the driver for this ZONE_DEVICE
4651 if (altmap && start_pfn == altmap->base_pfn)
4652 start_pfn += altmap->reserve;
4654 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4656 * There can be holes in boot-time mem_map[]s handed to this
4657 * function. They do not exist on hotplugged memory.
4659 if (context != MEMMAP_EARLY)
4662 if (!early_pfn_valid(pfn))
4664 if (!early_pfn_in_nid(pfn, nid))
4666 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4669 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4671 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4672 * from zone_movable_pfn[nid] to end of each node should be
4673 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4675 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4676 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4680 * Check given memblock attribute by firmware which can affect
4681 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4682 * mirrored, it's an overlapped memmap init. skip it.
4684 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4685 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4686 for_each_memblock(memory, tmp)
4687 if (pfn < memblock_region_memory_end_pfn(tmp))
4691 if (pfn >= memblock_region_memory_base_pfn(r) &&
4692 memblock_is_mirror(r)) {
4693 /* already initialized as NORMAL */
4694 pfn = memblock_region_memory_end_pfn(r);
4702 * Mark the block movable so that blocks are reserved for
4703 * movable at startup. This will force kernel allocations
4704 * to reserve their blocks rather than leaking throughout
4705 * the address space during boot when many long-lived
4706 * kernel allocations are made.
4708 * bitmap is created for zone's valid pfn range. but memmap
4709 * can be created for invalid pages (for alignment)
4710 * check here not to call set_pageblock_migratetype() against
4713 if (!(pfn & (pageblock_nr_pages - 1))) {
4714 struct page *page = pfn_to_page(pfn);
4716 __init_single_page(page, pfn, zone, nid);
4717 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4719 __init_single_pfn(pfn, zone, nid);
4724 static void __meminit zone_init_free_lists(struct zone *zone)
4726 unsigned int order, t;
4727 for_each_migratetype_order(order, t) {
4728 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4729 zone->free_area[order].nr_free = 0;
4733 #ifndef __HAVE_ARCH_MEMMAP_INIT
4734 #define memmap_init(size, nid, zone, start_pfn) \
4735 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4738 static int zone_batchsize(struct zone *zone)
4744 * The per-cpu-pages pools are set to around 1000th of the
4745 * size of the zone. But no more than 1/2 of a meg.
4747 * OK, so we don't know how big the cache is. So guess.
4749 batch = zone->managed_pages / 1024;
4750 if (batch * PAGE_SIZE > 512 * 1024)
4751 batch = (512 * 1024) / PAGE_SIZE;
4752 batch /= 4; /* We effectively *= 4 below */
4757 * Clamp the batch to a 2^n - 1 value. Having a power
4758 * of 2 value was found to be more likely to have
4759 * suboptimal cache aliasing properties in some cases.
4761 * For example if 2 tasks are alternately allocating
4762 * batches of pages, one task can end up with a lot
4763 * of pages of one half of the possible page colors
4764 * and the other with pages of the other colors.
4766 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4771 /* The deferral and batching of frees should be suppressed under NOMMU
4774 * The problem is that NOMMU needs to be able to allocate large chunks
4775 * of contiguous memory as there's no hardware page translation to
4776 * assemble apparent contiguous memory from discontiguous pages.
4778 * Queueing large contiguous runs of pages for batching, however,
4779 * causes the pages to actually be freed in smaller chunks. As there
4780 * can be a significant delay between the individual batches being
4781 * recycled, this leads to the once large chunks of space being
4782 * fragmented and becoming unavailable for high-order allocations.
4789 * pcp->high and pcp->batch values are related and dependent on one another:
4790 * ->batch must never be higher then ->high.
4791 * The following function updates them in a safe manner without read side
4794 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4795 * those fields changing asynchronously (acording the the above rule).
4797 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4798 * outside of boot time (or some other assurance that no concurrent updaters
4801 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4802 unsigned long batch)
4804 /* start with a fail safe value for batch */
4808 /* Update high, then batch, in order */
4815 /* a companion to pageset_set_high() */
4816 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4818 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4821 static void pageset_init(struct per_cpu_pageset *p)
4823 struct per_cpu_pages *pcp;
4826 memset(p, 0, sizeof(*p));
4830 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4831 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4834 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4837 pageset_set_batch(p, batch);
4841 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4842 * to the value high for the pageset p.
4844 static void pageset_set_high(struct per_cpu_pageset *p,
4847 unsigned long batch = max(1UL, high / 4);
4848 if ((high / 4) > (PAGE_SHIFT * 8))
4849 batch = PAGE_SHIFT * 8;
4851 pageset_update(&p->pcp, high, batch);
4854 static void pageset_set_high_and_batch(struct zone *zone,
4855 struct per_cpu_pageset *pcp)
4857 if (percpu_pagelist_fraction)
4858 pageset_set_high(pcp,
4859 (zone->managed_pages /
4860 percpu_pagelist_fraction));
4862 pageset_set_batch(pcp, zone_batchsize(zone));
4865 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4867 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4870 pageset_set_high_and_batch(zone, pcp);
4873 static void __meminit setup_zone_pageset(struct zone *zone)
4876 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4877 for_each_possible_cpu(cpu)
4878 zone_pageset_init(zone, cpu);
4882 * Allocate per cpu pagesets and initialize them.
4883 * Before this call only boot pagesets were available.
4885 void __init setup_per_cpu_pageset(void)
4889 for_each_populated_zone(zone)
4890 setup_zone_pageset(zone);
4893 static noinline __init_refok
4894 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4900 * The per-page waitqueue mechanism uses hashed waitqueues
4903 zone->wait_table_hash_nr_entries =
4904 wait_table_hash_nr_entries(zone_size_pages);
4905 zone->wait_table_bits =
4906 wait_table_bits(zone->wait_table_hash_nr_entries);
4907 alloc_size = zone->wait_table_hash_nr_entries
4908 * sizeof(wait_queue_head_t);
4910 if (!slab_is_available()) {
4911 zone->wait_table = (wait_queue_head_t *)
4912 memblock_virt_alloc_node_nopanic(
4913 alloc_size, zone->zone_pgdat->node_id);
4916 * This case means that a zone whose size was 0 gets new memory
4917 * via memory hot-add.
4918 * But it may be the case that a new node was hot-added. In
4919 * this case vmalloc() will not be able to use this new node's
4920 * memory - this wait_table must be initialized to use this new
4921 * node itself as well.
4922 * To use this new node's memory, further consideration will be
4925 zone->wait_table = vmalloc(alloc_size);
4927 if (!zone->wait_table)
4930 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4931 init_waitqueue_head(zone->wait_table + i);
4936 static __meminit void zone_pcp_init(struct zone *zone)
4939 * per cpu subsystem is not up at this point. The following code
4940 * relies on the ability of the linker to provide the
4941 * offset of a (static) per cpu variable into the per cpu area.
4943 zone->pageset = &boot_pageset;
4945 if (populated_zone(zone))
4946 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4947 zone->name, zone->present_pages,
4948 zone_batchsize(zone));
4951 int __meminit init_currently_empty_zone(struct zone *zone,
4952 unsigned long zone_start_pfn,
4955 struct pglist_data *pgdat = zone->zone_pgdat;
4957 ret = zone_wait_table_init(zone, size);
4960 pgdat->nr_zones = zone_idx(zone) + 1;
4962 zone->zone_start_pfn = zone_start_pfn;
4964 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4965 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4967 (unsigned long)zone_idx(zone),
4968 zone_start_pfn, (zone_start_pfn + size));
4970 zone_init_free_lists(zone);
4975 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4976 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4979 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4981 int __meminit __early_pfn_to_nid(unsigned long pfn,
4982 struct mminit_pfnnid_cache *state)
4984 unsigned long start_pfn, end_pfn;
4987 if (state->last_start <= pfn && pfn < state->last_end)
4988 return state->last_nid;
4990 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4992 state->last_start = start_pfn;
4993 state->last_end = end_pfn;
4994 state->last_nid = nid;
4999 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5002 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5003 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5004 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5006 * If an architecture guarantees that all ranges registered contain no holes
5007 * and may be freed, this this function may be used instead of calling
5008 * memblock_free_early_nid() manually.
5010 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5012 unsigned long start_pfn, end_pfn;
5015 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5016 start_pfn = min(start_pfn, max_low_pfn);
5017 end_pfn = min(end_pfn, max_low_pfn);
5019 if (start_pfn < end_pfn)
5020 memblock_free_early_nid(PFN_PHYS(start_pfn),
5021 (end_pfn - start_pfn) << PAGE_SHIFT,
5027 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5028 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5030 * If an architecture guarantees that all ranges registered contain no holes and may
5031 * be freed, this function may be used instead of calling memory_present() manually.
5033 void __init sparse_memory_present_with_active_regions(int nid)
5035 unsigned long start_pfn, end_pfn;
5038 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5039 memory_present(this_nid, start_pfn, end_pfn);
5043 * get_pfn_range_for_nid - Return the start and end page frames for a node
5044 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5045 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5046 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5048 * It returns the start and end page frame of a node based on information
5049 * provided by memblock_set_node(). If called for a node
5050 * with no available memory, a warning is printed and the start and end
5053 void __meminit get_pfn_range_for_nid(unsigned int nid,
5054 unsigned long *start_pfn, unsigned long *end_pfn)
5056 unsigned long this_start_pfn, this_end_pfn;
5062 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5063 *start_pfn = min(*start_pfn, this_start_pfn);
5064 *end_pfn = max(*end_pfn, this_end_pfn);
5067 if (*start_pfn == -1UL)
5072 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5073 * assumption is made that zones within a node are ordered in monotonic
5074 * increasing memory addresses so that the "highest" populated zone is used
5076 static void __init find_usable_zone_for_movable(void)
5079 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5080 if (zone_index == ZONE_MOVABLE)
5083 if (arch_zone_highest_possible_pfn[zone_index] >
5084 arch_zone_lowest_possible_pfn[zone_index])
5088 VM_BUG_ON(zone_index == -1);
5089 movable_zone = zone_index;
5093 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5094 * because it is sized independent of architecture. Unlike the other zones,
5095 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5096 * in each node depending on the size of each node and how evenly kernelcore
5097 * is distributed. This helper function adjusts the zone ranges
5098 * provided by the architecture for a given node by using the end of the
5099 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5100 * zones within a node are in order of monotonic increases memory addresses
5102 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5103 unsigned long zone_type,
5104 unsigned long node_start_pfn,
5105 unsigned long node_end_pfn,
5106 unsigned long *zone_start_pfn,
5107 unsigned long *zone_end_pfn)
5109 /* Only adjust if ZONE_MOVABLE is on this node */
5110 if (zone_movable_pfn[nid]) {
5111 /* Size ZONE_MOVABLE */
5112 if (zone_type == ZONE_MOVABLE) {
5113 *zone_start_pfn = zone_movable_pfn[nid];
5114 *zone_end_pfn = min(node_end_pfn,
5115 arch_zone_highest_possible_pfn[movable_zone]);
5117 /* Check if this whole range is within ZONE_MOVABLE */
5118 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5119 *zone_start_pfn = *zone_end_pfn;
5124 * Return the number of pages a zone spans in a node, including holes
5125 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5127 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5128 unsigned long zone_type,
5129 unsigned long node_start_pfn,
5130 unsigned long node_end_pfn,
5131 unsigned long *zone_start_pfn,
5132 unsigned long *zone_end_pfn,
5133 unsigned long *ignored)
5135 /* When hotadd a new node from cpu_up(), the node should be empty */
5136 if (!node_start_pfn && !node_end_pfn)
5139 /* Get the start and end of the zone */
5140 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5141 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5142 adjust_zone_range_for_zone_movable(nid, zone_type,
5143 node_start_pfn, node_end_pfn,
5144 zone_start_pfn, zone_end_pfn);
5146 /* Check that this node has pages within the zone's required range */
5147 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5150 /* Move the zone boundaries inside the node if necessary */
5151 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5152 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5154 /* Return the spanned pages */
5155 return *zone_end_pfn - *zone_start_pfn;
5159 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5160 * then all holes in the requested range will be accounted for.
5162 unsigned long __meminit __absent_pages_in_range(int nid,
5163 unsigned long range_start_pfn,
5164 unsigned long range_end_pfn)
5166 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5167 unsigned long start_pfn, end_pfn;
5170 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5171 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5172 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5173 nr_absent -= end_pfn - start_pfn;
5179 * absent_pages_in_range - Return number of page frames in holes within a range
5180 * @start_pfn: The start PFN to start searching for holes
5181 * @end_pfn: The end PFN to stop searching for holes
5183 * It returns the number of pages frames in memory holes within a range.
5185 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5186 unsigned long end_pfn)
5188 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5191 /* Return the number of page frames in holes in a zone on a node */
5192 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5193 unsigned long zone_type,
5194 unsigned long node_start_pfn,
5195 unsigned long node_end_pfn,
5196 unsigned long *ignored)
5198 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5199 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5200 unsigned long zone_start_pfn, zone_end_pfn;
5201 unsigned long nr_absent;
5203 /* When hotadd a new node from cpu_up(), the node should be empty */
5204 if (!node_start_pfn && !node_end_pfn)
5207 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5208 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5210 adjust_zone_range_for_zone_movable(nid, zone_type,
5211 node_start_pfn, node_end_pfn,
5212 &zone_start_pfn, &zone_end_pfn);
5213 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5216 * ZONE_MOVABLE handling.
5217 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5220 if (zone_movable_pfn[nid]) {
5221 if (mirrored_kernelcore) {
5222 unsigned long start_pfn, end_pfn;
5223 struct memblock_region *r;
5225 for_each_memblock(memory, r) {
5226 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5227 zone_start_pfn, zone_end_pfn);
5228 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5229 zone_start_pfn, zone_end_pfn);
5231 if (zone_type == ZONE_MOVABLE &&
5232 memblock_is_mirror(r))
5233 nr_absent += end_pfn - start_pfn;
5235 if (zone_type == ZONE_NORMAL &&
5236 !memblock_is_mirror(r))
5237 nr_absent += end_pfn - start_pfn;
5240 if (zone_type == ZONE_NORMAL)
5241 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5248 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5249 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5250 unsigned long zone_type,
5251 unsigned long node_start_pfn,
5252 unsigned long node_end_pfn,
5253 unsigned long *zone_start_pfn,
5254 unsigned long *zone_end_pfn,
5255 unsigned long *zones_size)
5259 *zone_start_pfn = node_start_pfn;
5260 for (zone = 0; zone < zone_type; zone++)
5261 *zone_start_pfn += zones_size[zone];
5263 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5265 return zones_size[zone_type];
5268 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5269 unsigned long zone_type,
5270 unsigned long node_start_pfn,
5271 unsigned long node_end_pfn,
5272 unsigned long *zholes_size)
5277 return zholes_size[zone_type];
5280 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5282 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5283 unsigned long node_start_pfn,
5284 unsigned long node_end_pfn,
5285 unsigned long *zones_size,
5286 unsigned long *zholes_size)
5288 unsigned long realtotalpages = 0, totalpages = 0;
5291 for (i = 0; i < MAX_NR_ZONES; i++) {
5292 struct zone *zone = pgdat->node_zones + i;
5293 unsigned long zone_start_pfn, zone_end_pfn;
5294 unsigned long size, real_size;
5296 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5302 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5303 node_start_pfn, node_end_pfn,
5306 zone->zone_start_pfn = zone_start_pfn;
5308 zone->zone_start_pfn = 0;
5309 zone->spanned_pages = size;
5310 zone->present_pages = real_size;
5313 realtotalpages += real_size;
5316 pgdat->node_spanned_pages = totalpages;
5317 pgdat->node_present_pages = realtotalpages;
5318 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5322 #ifndef CONFIG_SPARSEMEM
5324 * Calculate the size of the zone->blockflags rounded to an unsigned long
5325 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5326 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5327 * round what is now in bits to nearest long in bits, then return it in
5330 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5332 unsigned long usemapsize;
5334 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5335 usemapsize = roundup(zonesize, pageblock_nr_pages);
5336 usemapsize = usemapsize >> pageblock_order;
5337 usemapsize *= NR_PAGEBLOCK_BITS;
5338 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5340 return usemapsize / 8;
5343 static void __init setup_usemap(struct pglist_data *pgdat,
5345 unsigned long zone_start_pfn,
5346 unsigned long zonesize)
5348 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5349 zone->pageblock_flags = NULL;
5351 zone->pageblock_flags =
5352 memblock_virt_alloc_node_nopanic(usemapsize,
5356 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5357 unsigned long zone_start_pfn, unsigned long zonesize) {}
5358 #endif /* CONFIG_SPARSEMEM */
5360 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5362 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5363 void __paginginit set_pageblock_order(void)
5367 /* Check that pageblock_nr_pages has not already been setup */
5368 if (pageblock_order)
5371 if (HPAGE_SHIFT > PAGE_SHIFT)
5372 order = HUGETLB_PAGE_ORDER;
5374 order = MAX_ORDER - 1;
5377 * Assume the largest contiguous order of interest is a huge page.
5378 * This value may be variable depending on boot parameters on IA64 and
5381 pageblock_order = order;
5383 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5386 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5387 * is unused as pageblock_order is set at compile-time. See
5388 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5391 void __paginginit set_pageblock_order(void)
5395 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5397 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5398 unsigned long present_pages)
5400 unsigned long pages = spanned_pages;
5403 * Provide a more accurate estimation if there are holes within
5404 * the zone and SPARSEMEM is in use. If there are holes within the
5405 * zone, each populated memory region may cost us one or two extra
5406 * memmap pages due to alignment because memmap pages for each
5407 * populated regions may not naturally algined on page boundary.
5408 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5410 if (spanned_pages > present_pages + (present_pages >> 4) &&
5411 IS_ENABLED(CONFIG_SPARSEMEM))
5412 pages = present_pages;
5414 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5418 * Set up the zone data structures:
5419 * - mark all pages reserved
5420 * - mark all memory queues empty
5421 * - clear the memory bitmaps
5423 * NOTE: pgdat should get zeroed by caller.
5425 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5428 int nid = pgdat->node_id;
5431 pgdat_resize_init(pgdat);
5432 #ifdef CONFIG_NUMA_BALANCING
5433 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5434 pgdat->numabalancing_migrate_nr_pages = 0;
5435 pgdat->numabalancing_migrate_next_window = jiffies;
5437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5438 spin_lock_init(&pgdat->split_queue_lock);
5439 INIT_LIST_HEAD(&pgdat->split_queue);
5440 pgdat->split_queue_len = 0;
5442 init_waitqueue_head(&pgdat->kswapd_wait);
5443 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5444 #ifdef CONFIG_COMPACTION
5445 init_waitqueue_head(&pgdat->kcompactd_wait);
5447 pgdat_page_ext_init(pgdat);
5449 for (j = 0; j < MAX_NR_ZONES; j++) {
5450 struct zone *zone = pgdat->node_zones + j;
5451 unsigned long size, realsize, freesize, memmap_pages;
5452 unsigned long zone_start_pfn = zone->zone_start_pfn;
5454 size = zone->spanned_pages;
5455 realsize = freesize = zone->present_pages;
5458 * Adjust freesize so that it accounts for how much memory
5459 * is used by this zone for memmap. This affects the watermark
5460 * and per-cpu initialisations
5462 memmap_pages = calc_memmap_size(size, realsize);
5463 if (!is_highmem_idx(j)) {
5464 if (freesize >= memmap_pages) {
5465 freesize -= memmap_pages;
5468 " %s zone: %lu pages used for memmap\n",
5469 zone_names[j], memmap_pages);
5471 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5472 zone_names[j], memmap_pages, freesize);
5475 /* Account for reserved pages */
5476 if (j == 0 && freesize > dma_reserve) {
5477 freesize -= dma_reserve;
5478 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5479 zone_names[0], dma_reserve);
5482 if (!is_highmem_idx(j))
5483 nr_kernel_pages += freesize;
5484 /* Charge for highmem memmap if there are enough kernel pages */
5485 else if (nr_kernel_pages > memmap_pages * 2)
5486 nr_kernel_pages -= memmap_pages;
5487 nr_all_pages += freesize;
5490 * Set an approximate value for lowmem here, it will be adjusted
5491 * when the bootmem allocator frees pages into the buddy system.
5492 * And all highmem pages will be managed by the buddy system.
5494 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5497 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5499 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5501 zone->name = zone_names[j];
5502 spin_lock_init(&zone->lock);
5503 spin_lock_init(&zone->lru_lock);
5504 zone_seqlock_init(zone);
5505 zone->zone_pgdat = pgdat;
5506 zone_pcp_init(zone);
5508 /* For bootup, initialized properly in watermark setup */
5509 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5511 lruvec_init(&zone->lruvec);
5515 set_pageblock_order();
5516 setup_usemap(pgdat, zone, zone_start_pfn, size);
5517 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5519 memmap_init(size, nid, j, zone_start_pfn);
5523 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5525 unsigned long __maybe_unused start = 0;
5526 unsigned long __maybe_unused offset = 0;
5528 /* Skip empty nodes */
5529 if (!pgdat->node_spanned_pages)
5532 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5533 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5534 offset = pgdat->node_start_pfn - start;
5535 /* ia64 gets its own node_mem_map, before this, without bootmem */
5536 if (!pgdat->node_mem_map) {
5537 unsigned long size, end;
5541 * The zone's endpoints aren't required to be MAX_ORDER
5542 * aligned but the node_mem_map endpoints must be in order
5543 * for the buddy allocator to function correctly.
5545 end = pgdat_end_pfn(pgdat);
5546 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5547 size = (end - start) * sizeof(struct page);
5548 map = alloc_remap(pgdat->node_id, size);
5550 map = memblock_virt_alloc_node_nopanic(size,
5552 pgdat->node_mem_map = map + offset;
5554 #ifndef CONFIG_NEED_MULTIPLE_NODES
5556 * With no DISCONTIG, the global mem_map is just set as node 0's
5558 if (pgdat == NODE_DATA(0)) {
5559 mem_map = NODE_DATA(0)->node_mem_map;
5560 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5561 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5563 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5566 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5569 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5570 unsigned long node_start_pfn, unsigned long *zholes_size)
5572 pg_data_t *pgdat = NODE_DATA(nid);
5573 unsigned long start_pfn = 0;
5574 unsigned long end_pfn = 0;
5576 /* pg_data_t should be reset to zero when it's allocated */
5577 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5579 reset_deferred_meminit(pgdat);
5580 pgdat->node_id = nid;
5581 pgdat->node_start_pfn = node_start_pfn;
5582 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5583 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5584 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5585 (u64)start_pfn << PAGE_SHIFT,
5586 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5588 start_pfn = node_start_pfn;
5590 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5591 zones_size, zholes_size);
5593 alloc_node_mem_map(pgdat);
5594 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5595 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5596 nid, (unsigned long)pgdat,
5597 (unsigned long)pgdat->node_mem_map);
5600 free_area_init_core(pgdat);
5603 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5605 #if MAX_NUMNODES > 1
5607 * Figure out the number of possible node ids.
5609 void __init setup_nr_node_ids(void)
5611 unsigned int highest;
5613 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5614 nr_node_ids = highest + 1;
5619 * node_map_pfn_alignment - determine the maximum internode alignment
5621 * This function should be called after node map is populated and sorted.
5622 * It calculates the maximum power of two alignment which can distinguish
5625 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5626 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5627 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5628 * shifted, 1GiB is enough and this function will indicate so.
5630 * This is used to test whether pfn -> nid mapping of the chosen memory
5631 * model has fine enough granularity to avoid incorrect mapping for the
5632 * populated node map.
5634 * Returns the determined alignment in pfn's. 0 if there is no alignment
5635 * requirement (single node).
5637 unsigned long __init node_map_pfn_alignment(void)
5639 unsigned long accl_mask = 0, last_end = 0;
5640 unsigned long start, end, mask;
5644 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5645 if (!start || last_nid < 0 || last_nid == nid) {
5652 * Start with a mask granular enough to pin-point to the
5653 * start pfn and tick off bits one-by-one until it becomes
5654 * too coarse to separate the current node from the last.
5656 mask = ~((1 << __ffs(start)) - 1);
5657 while (mask && last_end <= (start & (mask << 1)))
5660 /* accumulate all internode masks */
5664 /* convert mask to number of pages */
5665 return ~accl_mask + 1;
5668 /* Find the lowest pfn for a node */
5669 static unsigned long __init find_min_pfn_for_node(int nid)
5671 unsigned long min_pfn = ULONG_MAX;
5672 unsigned long start_pfn;
5675 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5676 min_pfn = min(min_pfn, start_pfn);
5678 if (min_pfn == ULONG_MAX) {
5679 pr_warn("Could not find start_pfn for node %d\n", nid);
5687 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5689 * It returns the minimum PFN based on information provided via
5690 * memblock_set_node().
5692 unsigned long __init find_min_pfn_with_active_regions(void)
5694 return find_min_pfn_for_node(MAX_NUMNODES);
5698 * early_calculate_totalpages()
5699 * Sum pages in active regions for movable zone.
5700 * Populate N_MEMORY for calculating usable_nodes.
5702 static unsigned long __init early_calculate_totalpages(void)
5704 unsigned long totalpages = 0;
5705 unsigned long start_pfn, end_pfn;
5708 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5709 unsigned long pages = end_pfn - start_pfn;
5711 totalpages += pages;
5713 node_set_state(nid, N_MEMORY);
5719 * Find the PFN the Movable zone begins in each node. Kernel memory
5720 * is spread evenly between nodes as long as the nodes have enough
5721 * memory. When they don't, some nodes will have more kernelcore than
5724 static void __init find_zone_movable_pfns_for_nodes(void)
5727 unsigned long usable_startpfn;
5728 unsigned long kernelcore_node, kernelcore_remaining;
5729 /* save the state before borrow the nodemask */
5730 nodemask_t saved_node_state = node_states[N_MEMORY];
5731 unsigned long totalpages = early_calculate_totalpages();
5732 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5733 struct memblock_region *r;
5735 /* Need to find movable_zone earlier when movable_node is specified. */
5736 find_usable_zone_for_movable();
5739 * If movable_node is specified, ignore kernelcore and movablecore
5742 if (movable_node_is_enabled()) {
5743 for_each_memblock(memory, r) {
5744 if (!memblock_is_hotpluggable(r))
5749 usable_startpfn = PFN_DOWN(r->base);
5750 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5751 min(usable_startpfn, zone_movable_pfn[nid]) :
5759 * If kernelcore=mirror is specified, ignore movablecore option
5761 if (mirrored_kernelcore) {
5762 bool mem_below_4gb_not_mirrored = false;
5764 for_each_memblock(memory, r) {
5765 if (memblock_is_mirror(r))
5770 usable_startpfn = memblock_region_memory_base_pfn(r);
5772 if (usable_startpfn < 0x100000) {
5773 mem_below_4gb_not_mirrored = true;
5777 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5778 min(usable_startpfn, zone_movable_pfn[nid]) :
5782 if (mem_below_4gb_not_mirrored)
5783 pr_warn("This configuration results in unmirrored kernel memory.");
5789 * If movablecore=nn[KMG] was specified, calculate what size of
5790 * kernelcore that corresponds so that memory usable for
5791 * any allocation type is evenly spread. If both kernelcore
5792 * and movablecore are specified, then the value of kernelcore
5793 * will be used for required_kernelcore if it's greater than
5794 * what movablecore would have allowed.
5796 if (required_movablecore) {
5797 unsigned long corepages;
5800 * Round-up so that ZONE_MOVABLE is at least as large as what
5801 * was requested by the user
5803 required_movablecore =
5804 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5805 required_movablecore = min(totalpages, required_movablecore);
5806 corepages = totalpages - required_movablecore;
5808 required_kernelcore = max(required_kernelcore, corepages);
5812 * If kernelcore was not specified or kernelcore size is larger
5813 * than totalpages, there is no ZONE_MOVABLE.
5815 if (!required_kernelcore || required_kernelcore >= totalpages)
5818 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5819 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5822 /* Spread kernelcore memory as evenly as possible throughout nodes */
5823 kernelcore_node = required_kernelcore / usable_nodes;
5824 for_each_node_state(nid, N_MEMORY) {
5825 unsigned long start_pfn, end_pfn;
5828 * Recalculate kernelcore_node if the division per node
5829 * now exceeds what is necessary to satisfy the requested
5830 * amount of memory for the kernel
5832 if (required_kernelcore < kernelcore_node)
5833 kernelcore_node = required_kernelcore / usable_nodes;
5836 * As the map is walked, we track how much memory is usable
5837 * by the kernel using kernelcore_remaining. When it is
5838 * 0, the rest of the node is usable by ZONE_MOVABLE
5840 kernelcore_remaining = kernelcore_node;
5842 /* Go through each range of PFNs within this node */
5843 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5844 unsigned long size_pages;
5846 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5847 if (start_pfn >= end_pfn)
5850 /* Account for what is only usable for kernelcore */
5851 if (start_pfn < usable_startpfn) {
5852 unsigned long kernel_pages;
5853 kernel_pages = min(end_pfn, usable_startpfn)
5856 kernelcore_remaining -= min(kernel_pages,
5857 kernelcore_remaining);
5858 required_kernelcore -= min(kernel_pages,
5859 required_kernelcore);
5861 /* Continue if range is now fully accounted */
5862 if (end_pfn <= usable_startpfn) {
5865 * Push zone_movable_pfn to the end so
5866 * that if we have to rebalance
5867 * kernelcore across nodes, we will
5868 * not double account here
5870 zone_movable_pfn[nid] = end_pfn;
5873 start_pfn = usable_startpfn;
5877 * The usable PFN range for ZONE_MOVABLE is from
5878 * start_pfn->end_pfn. Calculate size_pages as the
5879 * number of pages used as kernelcore
5881 size_pages = end_pfn - start_pfn;
5882 if (size_pages > kernelcore_remaining)
5883 size_pages = kernelcore_remaining;
5884 zone_movable_pfn[nid] = start_pfn + size_pages;
5887 * Some kernelcore has been met, update counts and
5888 * break if the kernelcore for this node has been
5891 required_kernelcore -= min(required_kernelcore,
5893 kernelcore_remaining -= size_pages;
5894 if (!kernelcore_remaining)
5900 * If there is still required_kernelcore, we do another pass with one
5901 * less node in the count. This will push zone_movable_pfn[nid] further
5902 * along on the nodes that still have memory until kernelcore is
5906 if (usable_nodes && required_kernelcore > usable_nodes)
5910 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5911 for (nid = 0; nid < MAX_NUMNODES; nid++)
5912 zone_movable_pfn[nid] =
5913 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5916 /* restore the node_state */
5917 node_states[N_MEMORY] = saved_node_state;
5920 /* Any regular or high memory on that node ? */
5921 static void check_for_memory(pg_data_t *pgdat, int nid)
5923 enum zone_type zone_type;
5925 if (N_MEMORY == N_NORMAL_MEMORY)
5928 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5929 struct zone *zone = &pgdat->node_zones[zone_type];
5930 if (populated_zone(zone)) {
5931 node_set_state(nid, N_HIGH_MEMORY);
5932 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5933 zone_type <= ZONE_NORMAL)
5934 node_set_state(nid, N_NORMAL_MEMORY);
5941 * free_area_init_nodes - Initialise all pg_data_t and zone data
5942 * @max_zone_pfn: an array of max PFNs for each zone
5944 * This will call free_area_init_node() for each active node in the system.
5945 * Using the page ranges provided by memblock_set_node(), the size of each
5946 * zone in each node and their holes is calculated. If the maximum PFN
5947 * between two adjacent zones match, it is assumed that the zone is empty.
5948 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5949 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5950 * starts where the previous one ended. For example, ZONE_DMA32 starts
5951 * at arch_max_dma_pfn.
5953 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5955 unsigned long start_pfn, end_pfn;
5958 /* Record where the zone boundaries are */
5959 memset(arch_zone_lowest_possible_pfn, 0,
5960 sizeof(arch_zone_lowest_possible_pfn));
5961 memset(arch_zone_highest_possible_pfn, 0,
5962 sizeof(arch_zone_highest_possible_pfn));
5963 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5964 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5965 for (i = 1; i < MAX_NR_ZONES; i++) {
5966 if (i == ZONE_MOVABLE)
5968 arch_zone_lowest_possible_pfn[i] =
5969 arch_zone_highest_possible_pfn[i-1];
5970 arch_zone_highest_possible_pfn[i] =
5971 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5973 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5974 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5976 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5977 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5978 find_zone_movable_pfns_for_nodes();
5980 /* Print out the zone ranges */
5981 pr_info("Zone ranges:\n");
5982 for (i = 0; i < MAX_NR_ZONES; i++) {
5983 if (i == ZONE_MOVABLE)
5985 pr_info(" %-8s ", zone_names[i]);
5986 if (arch_zone_lowest_possible_pfn[i] ==
5987 arch_zone_highest_possible_pfn[i])
5990 pr_cont("[mem %#018Lx-%#018Lx]\n",
5991 (u64)arch_zone_lowest_possible_pfn[i]
5993 ((u64)arch_zone_highest_possible_pfn[i]
5994 << PAGE_SHIFT) - 1);
5997 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5998 pr_info("Movable zone start for each node\n");
5999 for (i = 0; i < MAX_NUMNODES; i++) {
6000 if (zone_movable_pfn[i])
6001 pr_info(" Node %d: %#018Lx\n", i,
6002 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6005 /* Print out the early node map */
6006 pr_info("Early memory node ranges\n");
6007 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6008 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6009 (u64)start_pfn << PAGE_SHIFT,
6010 ((u64)end_pfn << PAGE_SHIFT) - 1);
6012 /* Initialise every node */
6013 mminit_verify_pageflags_layout();
6014 setup_nr_node_ids();
6015 for_each_online_node(nid) {
6016 pg_data_t *pgdat = NODE_DATA(nid);
6017 free_area_init_node(nid, NULL,
6018 find_min_pfn_for_node(nid), NULL);
6020 /* Any memory on that node */
6021 if (pgdat->node_present_pages)
6022 node_set_state(nid, N_MEMORY);
6023 check_for_memory(pgdat, nid);
6027 static int __init cmdline_parse_core(char *p, unsigned long *core)
6029 unsigned long long coremem;
6033 coremem = memparse(p, &p);
6034 *core = coremem >> PAGE_SHIFT;
6036 /* Paranoid check that UL is enough for the coremem value */
6037 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6043 * kernelcore=size sets the amount of memory for use for allocations that
6044 * cannot be reclaimed or migrated.
6046 static int __init cmdline_parse_kernelcore(char *p)
6048 /* parse kernelcore=mirror */
6049 if (parse_option_str(p, "mirror")) {
6050 mirrored_kernelcore = true;
6054 return cmdline_parse_core(p, &required_kernelcore);
6058 * movablecore=size sets the amount of memory for use for allocations that
6059 * can be reclaimed or migrated.
6061 static int __init cmdline_parse_movablecore(char *p)
6063 return cmdline_parse_core(p, &required_movablecore);
6066 early_param("kernelcore", cmdline_parse_kernelcore);
6067 early_param("movablecore", cmdline_parse_movablecore);
6069 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6071 void adjust_managed_page_count(struct page *page, long count)
6073 spin_lock(&managed_page_count_lock);
6074 page_zone(page)->managed_pages += count;
6075 totalram_pages += count;
6076 #ifdef CONFIG_HIGHMEM
6077 if (PageHighMem(page))
6078 totalhigh_pages += count;
6080 spin_unlock(&managed_page_count_lock);
6082 EXPORT_SYMBOL(adjust_managed_page_count);
6084 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6087 unsigned long pages = 0;
6089 start = (void *)PAGE_ALIGN((unsigned long)start);
6090 end = (void *)((unsigned long)end & PAGE_MASK);
6091 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6092 if ((unsigned int)poison <= 0xFF)
6093 memset(pos, poison, PAGE_SIZE);
6094 free_reserved_page(virt_to_page(pos));
6098 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6099 s, pages << (PAGE_SHIFT - 10), start, end);
6103 EXPORT_SYMBOL(free_reserved_area);
6105 #ifdef CONFIG_HIGHMEM
6106 void free_highmem_page(struct page *page)
6108 __free_reserved_page(page);
6110 page_zone(page)->managed_pages++;
6116 void __init mem_init_print_info(const char *str)
6118 unsigned long physpages, codesize, datasize, rosize, bss_size;
6119 unsigned long init_code_size, init_data_size;
6121 physpages = get_num_physpages();
6122 codesize = _etext - _stext;
6123 datasize = _edata - _sdata;
6124 rosize = __end_rodata - __start_rodata;
6125 bss_size = __bss_stop - __bss_start;
6126 init_data_size = __init_end - __init_begin;
6127 init_code_size = _einittext - _sinittext;
6130 * Detect special cases and adjust section sizes accordingly:
6131 * 1) .init.* may be embedded into .data sections
6132 * 2) .init.text.* may be out of [__init_begin, __init_end],
6133 * please refer to arch/tile/kernel/vmlinux.lds.S.
6134 * 3) .rodata.* may be embedded into .text or .data sections.
6136 #define adj_init_size(start, end, size, pos, adj) \
6138 if (start <= pos && pos < end && size > adj) \
6142 adj_init_size(__init_begin, __init_end, init_data_size,
6143 _sinittext, init_code_size);
6144 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6145 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6146 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6147 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6149 #undef adj_init_size
6151 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6152 #ifdef CONFIG_HIGHMEM
6156 nr_free_pages() << (PAGE_SHIFT - 10),
6157 physpages << (PAGE_SHIFT - 10),
6158 codesize >> 10, datasize >> 10, rosize >> 10,
6159 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6160 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6161 totalcma_pages << (PAGE_SHIFT - 10),
6162 #ifdef CONFIG_HIGHMEM
6163 totalhigh_pages << (PAGE_SHIFT - 10),
6165 str ? ", " : "", str ? str : "");
6169 * set_dma_reserve - set the specified number of pages reserved in the first zone
6170 * @new_dma_reserve: The number of pages to mark reserved
6172 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6173 * In the DMA zone, a significant percentage may be consumed by kernel image
6174 * and other unfreeable allocations which can skew the watermarks badly. This
6175 * function may optionally be used to account for unfreeable pages in the
6176 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6177 * smaller per-cpu batchsize.
6179 void __init set_dma_reserve(unsigned long new_dma_reserve)
6181 dma_reserve = new_dma_reserve;
6184 void __init free_area_init(unsigned long *zones_size)
6186 free_area_init_node(0, zones_size,
6187 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6190 static int page_alloc_cpu_notify(struct notifier_block *self,
6191 unsigned long action, void *hcpu)
6193 int cpu = (unsigned long)hcpu;
6195 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6196 lru_add_drain_cpu(cpu);
6200 * Spill the event counters of the dead processor
6201 * into the current processors event counters.
6202 * This artificially elevates the count of the current
6205 vm_events_fold_cpu(cpu);
6208 * Zero the differential counters of the dead processor
6209 * so that the vm statistics are consistent.
6211 * This is only okay since the processor is dead and cannot
6212 * race with what we are doing.
6214 cpu_vm_stats_fold(cpu);
6219 void __init page_alloc_init(void)
6221 hotcpu_notifier(page_alloc_cpu_notify, 0);
6225 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6226 * or min_free_kbytes changes.
6228 static void calculate_totalreserve_pages(void)
6230 struct pglist_data *pgdat;
6231 unsigned long reserve_pages = 0;
6232 enum zone_type i, j;
6234 for_each_online_pgdat(pgdat) {
6235 for (i = 0; i < MAX_NR_ZONES; i++) {
6236 struct zone *zone = pgdat->node_zones + i;
6239 /* Find valid and maximum lowmem_reserve in the zone */
6240 for (j = i; j < MAX_NR_ZONES; j++) {
6241 if (zone->lowmem_reserve[j] > max)
6242 max = zone->lowmem_reserve[j];
6245 /* we treat the high watermark as reserved pages. */
6246 max += high_wmark_pages(zone);
6248 if (max > zone->managed_pages)
6249 max = zone->managed_pages;
6251 zone->totalreserve_pages = max;
6253 reserve_pages += max;
6256 totalreserve_pages = reserve_pages;
6260 * setup_per_zone_lowmem_reserve - called whenever
6261 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6262 * has a correct pages reserved value, so an adequate number of
6263 * pages are left in the zone after a successful __alloc_pages().
6265 static void setup_per_zone_lowmem_reserve(void)
6267 struct pglist_data *pgdat;
6268 enum zone_type j, idx;
6270 for_each_online_pgdat(pgdat) {
6271 for (j = 0; j < MAX_NR_ZONES; j++) {
6272 struct zone *zone = pgdat->node_zones + j;
6273 unsigned long managed_pages = zone->managed_pages;
6275 zone->lowmem_reserve[j] = 0;
6279 struct zone *lower_zone;
6283 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6284 sysctl_lowmem_reserve_ratio[idx] = 1;
6286 lower_zone = pgdat->node_zones + idx;
6287 lower_zone->lowmem_reserve[j] = managed_pages /
6288 sysctl_lowmem_reserve_ratio[idx];
6289 managed_pages += lower_zone->managed_pages;
6294 /* update totalreserve_pages */
6295 calculate_totalreserve_pages();
6298 static void __setup_per_zone_wmarks(void)
6300 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6301 unsigned long lowmem_pages = 0;
6303 unsigned long flags;
6305 /* Calculate total number of !ZONE_HIGHMEM pages */
6306 for_each_zone(zone) {
6307 if (!is_highmem(zone))
6308 lowmem_pages += zone->managed_pages;
6311 for_each_zone(zone) {
6314 spin_lock_irqsave(&zone->lock, flags);
6315 tmp = (u64)pages_min * zone->managed_pages;
6316 do_div(tmp, lowmem_pages);
6317 if (is_highmem(zone)) {
6319 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6320 * need highmem pages, so cap pages_min to a small
6323 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6324 * deltas control asynch page reclaim, and so should
6325 * not be capped for highmem.
6327 unsigned long min_pages;
6329 min_pages = zone->managed_pages / 1024;
6330 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6331 zone->watermark[WMARK_MIN] = min_pages;
6334 * If it's a lowmem zone, reserve a number of pages
6335 * proportionate to the zone's size.
6337 zone->watermark[WMARK_MIN] = tmp;
6341 * Set the kswapd watermarks distance according to the
6342 * scale factor in proportion to available memory, but
6343 * ensure a minimum size on small systems.
6345 tmp = max_t(u64, tmp >> 2,
6346 mult_frac(zone->managed_pages,
6347 watermark_scale_factor, 10000));
6349 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6350 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6352 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6353 high_wmark_pages(zone) - low_wmark_pages(zone) -
6354 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6356 spin_unlock_irqrestore(&zone->lock, flags);
6359 /* update totalreserve_pages */
6360 calculate_totalreserve_pages();
6364 * setup_per_zone_wmarks - called when min_free_kbytes changes
6365 * or when memory is hot-{added|removed}
6367 * Ensures that the watermark[min,low,high] values for each zone are set
6368 * correctly with respect to min_free_kbytes.
6370 void setup_per_zone_wmarks(void)
6372 mutex_lock(&zonelists_mutex);
6373 __setup_per_zone_wmarks();
6374 mutex_unlock(&zonelists_mutex);
6378 * The inactive anon list should be small enough that the VM never has to
6379 * do too much work, but large enough that each inactive page has a chance
6380 * to be referenced again before it is swapped out.
6382 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6383 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6384 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6385 * the anonymous pages are kept on the inactive list.
6388 * memory ratio inactive anon
6389 * -------------------------------------
6398 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6400 unsigned int gb, ratio;
6402 /* Zone size in gigabytes */
6403 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6405 ratio = int_sqrt(10 * gb);
6409 zone->inactive_ratio = ratio;
6412 static void __meminit setup_per_zone_inactive_ratio(void)
6417 calculate_zone_inactive_ratio(zone);
6421 * Initialise min_free_kbytes.
6423 * For small machines we want it small (128k min). For large machines
6424 * we want it large (64MB max). But it is not linear, because network
6425 * bandwidth does not increase linearly with machine size. We use
6427 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6428 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6444 int __meminit init_per_zone_wmark_min(void)
6446 unsigned long lowmem_kbytes;
6447 int new_min_free_kbytes;
6449 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6450 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6452 if (new_min_free_kbytes > user_min_free_kbytes) {
6453 min_free_kbytes = new_min_free_kbytes;
6454 if (min_free_kbytes < 128)
6455 min_free_kbytes = 128;
6456 if (min_free_kbytes > 65536)
6457 min_free_kbytes = 65536;
6459 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6460 new_min_free_kbytes, user_min_free_kbytes);
6462 setup_per_zone_wmarks();
6463 refresh_zone_stat_thresholds();
6464 setup_per_zone_lowmem_reserve();
6465 setup_per_zone_inactive_ratio();
6468 module_init(init_per_zone_wmark_min)
6471 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6472 * that we can call two helper functions whenever min_free_kbytes
6475 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6476 void __user *buffer, size_t *length, loff_t *ppos)
6480 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6485 user_min_free_kbytes = min_free_kbytes;
6486 setup_per_zone_wmarks();
6491 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6492 void __user *buffer, size_t *length, loff_t *ppos)
6496 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6501 setup_per_zone_wmarks();
6507 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6508 void __user *buffer, size_t *length, loff_t *ppos)
6513 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6518 zone->min_unmapped_pages = (zone->managed_pages *
6519 sysctl_min_unmapped_ratio) / 100;
6523 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6524 void __user *buffer, size_t *length, loff_t *ppos)
6529 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6534 zone->min_slab_pages = (zone->managed_pages *
6535 sysctl_min_slab_ratio) / 100;
6541 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6542 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6543 * whenever sysctl_lowmem_reserve_ratio changes.
6545 * The reserve ratio obviously has absolutely no relation with the
6546 * minimum watermarks. The lowmem reserve ratio can only make sense
6547 * if in function of the boot time zone sizes.
6549 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6550 void __user *buffer, size_t *length, loff_t *ppos)
6552 proc_dointvec_minmax(table, write, buffer, length, ppos);
6553 setup_per_zone_lowmem_reserve();
6558 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6559 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6560 * pagelist can have before it gets flushed back to buddy allocator.
6562 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6563 void __user *buffer, size_t *length, loff_t *ppos)
6566 int old_percpu_pagelist_fraction;
6569 mutex_lock(&pcp_batch_high_lock);
6570 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6572 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6573 if (!write || ret < 0)
6576 /* Sanity checking to avoid pcp imbalance */
6577 if (percpu_pagelist_fraction &&
6578 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6579 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6585 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6588 for_each_populated_zone(zone) {
6591 for_each_possible_cpu(cpu)
6592 pageset_set_high_and_batch(zone,
6593 per_cpu_ptr(zone->pageset, cpu));
6596 mutex_unlock(&pcp_batch_high_lock);
6601 int hashdist = HASHDIST_DEFAULT;
6603 static int __init set_hashdist(char *str)
6607 hashdist = simple_strtoul(str, &str, 0);
6610 __setup("hashdist=", set_hashdist);
6614 * allocate a large system hash table from bootmem
6615 * - it is assumed that the hash table must contain an exact power-of-2
6616 * quantity of entries
6617 * - limit is the number of hash buckets, not the total allocation size
6619 void *__init alloc_large_system_hash(const char *tablename,
6620 unsigned long bucketsize,
6621 unsigned long numentries,
6624 unsigned int *_hash_shift,
6625 unsigned int *_hash_mask,
6626 unsigned long low_limit,
6627 unsigned long high_limit)
6629 unsigned long long max = high_limit;
6630 unsigned long log2qty, size;
6633 /* allow the kernel cmdline to have a say */
6635 /* round applicable memory size up to nearest megabyte */
6636 numentries = nr_kernel_pages;
6638 /* It isn't necessary when PAGE_SIZE >= 1MB */
6639 if (PAGE_SHIFT < 20)
6640 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6642 /* limit to 1 bucket per 2^scale bytes of low memory */
6643 if (scale > PAGE_SHIFT)
6644 numentries >>= (scale - PAGE_SHIFT);
6646 numentries <<= (PAGE_SHIFT - scale);
6648 /* Make sure we've got at least a 0-order allocation.. */
6649 if (unlikely(flags & HASH_SMALL)) {
6650 /* Makes no sense without HASH_EARLY */
6651 WARN_ON(!(flags & HASH_EARLY));
6652 if (!(numentries >> *_hash_shift)) {
6653 numentries = 1UL << *_hash_shift;
6654 BUG_ON(!numentries);
6656 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6657 numentries = PAGE_SIZE / bucketsize;
6659 numentries = roundup_pow_of_two(numentries);
6661 /* limit allocation size to 1/16 total memory by default */
6663 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6664 do_div(max, bucketsize);
6666 max = min(max, 0x80000000ULL);
6668 if (numentries < low_limit)
6669 numentries = low_limit;
6670 if (numentries > max)
6673 log2qty = ilog2(numentries);
6676 size = bucketsize << log2qty;
6677 if (flags & HASH_EARLY)
6678 table = memblock_virt_alloc_nopanic(size, 0);
6680 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6683 * If bucketsize is not a power-of-two, we may free
6684 * some pages at the end of hash table which
6685 * alloc_pages_exact() automatically does
6687 if (get_order(size) < MAX_ORDER) {
6688 table = alloc_pages_exact(size, GFP_ATOMIC);
6689 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6692 } while (!table && size > PAGE_SIZE && --log2qty);
6695 panic("Failed to allocate %s hash table\n", tablename);
6697 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6698 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
6701 *_hash_shift = log2qty;
6703 *_hash_mask = (1 << log2qty) - 1;
6708 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6709 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6712 #ifdef CONFIG_SPARSEMEM
6713 return __pfn_to_section(pfn)->pageblock_flags;
6715 return zone->pageblock_flags;
6716 #endif /* CONFIG_SPARSEMEM */
6719 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6721 #ifdef CONFIG_SPARSEMEM
6722 pfn &= (PAGES_PER_SECTION-1);
6723 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6725 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6726 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6727 #endif /* CONFIG_SPARSEMEM */
6731 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6732 * @page: The page within the block of interest
6733 * @pfn: The target page frame number
6734 * @end_bitidx: The last bit of interest to retrieve
6735 * @mask: mask of bits that the caller is interested in
6737 * Return: pageblock_bits flags
6739 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6740 unsigned long end_bitidx,
6744 unsigned long *bitmap;
6745 unsigned long bitidx, word_bitidx;
6748 zone = page_zone(page);
6749 bitmap = get_pageblock_bitmap(zone, pfn);
6750 bitidx = pfn_to_bitidx(zone, pfn);
6751 word_bitidx = bitidx / BITS_PER_LONG;
6752 bitidx &= (BITS_PER_LONG-1);
6754 word = bitmap[word_bitidx];
6755 bitidx += end_bitidx;
6756 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6760 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6761 * @page: The page within the block of interest
6762 * @flags: The flags to set
6763 * @pfn: The target page frame number
6764 * @end_bitidx: The last bit of interest
6765 * @mask: mask of bits that the caller is interested in
6767 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6769 unsigned long end_bitidx,
6773 unsigned long *bitmap;
6774 unsigned long bitidx, word_bitidx;
6775 unsigned long old_word, word;
6777 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6779 zone = page_zone(page);
6780 bitmap = get_pageblock_bitmap(zone, pfn);
6781 bitidx = pfn_to_bitidx(zone, pfn);
6782 word_bitidx = bitidx / BITS_PER_LONG;
6783 bitidx &= (BITS_PER_LONG-1);
6785 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6787 bitidx += end_bitidx;
6788 mask <<= (BITS_PER_LONG - bitidx - 1);
6789 flags <<= (BITS_PER_LONG - bitidx - 1);
6791 word = READ_ONCE(bitmap[word_bitidx]);
6793 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6794 if (word == old_word)
6801 * This function checks whether pageblock includes unmovable pages or not.
6802 * If @count is not zero, it is okay to include less @count unmovable pages
6804 * PageLRU check without isolation or lru_lock could race so that
6805 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6806 * expect this function should be exact.
6808 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6809 bool skip_hwpoisoned_pages)
6811 unsigned long pfn, iter, found;
6815 * For avoiding noise data, lru_add_drain_all() should be called
6816 * If ZONE_MOVABLE, the zone never contains unmovable pages
6818 if (zone_idx(zone) == ZONE_MOVABLE)
6820 mt = get_pageblock_migratetype(page);
6821 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6824 pfn = page_to_pfn(page);
6825 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6826 unsigned long check = pfn + iter;
6828 if (!pfn_valid_within(check))
6831 page = pfn_to_page(check);
6834 * Hugepages are not in LRU lists, but they're movable.
6835 * We need not scan over tail pages bacause we don't
6836 * handle each tail page individually in migration.
6838 if (PageHuge(page)) {
6839 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6844 * We can't use page_count without pin a page
6845 * because another CPU can free compound page.
6846 * This check already skips compound tails of THP
6847 * because their page->_count is zero at all time.
6849 if (!page_ref_count(page)) {
6850 if (PageBuddy(page))
6851 iter += (1 << page_order(page)) - 1;
6856 * The HWPoisoned page may be not in buddy system, and
6857 * page_count() is not 0.
6859 if (skip_hwpoisoned_pages && PageHWPoison(page))
6865 * If there are RECLAIMABLE pages, we need to check
6866 * it. But now, memory offline itself doesn't call
6867 * shrink_node_slabs() and it still to be fixed.
6870 * If the page is not RAM, page_count()should be 0.
6871 * we don't need more check. This is an _used_ not-movable page.
6873 * The problematic thing here is PG_reserved pages. PG_reserved
6874 * is set to both of a memory hole page and a _used_ kernel
6883 bool is_pageblock_removable_nolock(struct page *page)
6889 * We have to be careful here because we are iterating over memory
6890 * sections which are not zone aware so we might end up outside of
6891 * the zone but still within the section.
6892 * We have to take care about the node as well. If the node is offline
6893 * its NODE_DATA will be NULL - see page_zone.
6895 if (!node_online(page_to_nid(page)))
6898 zone = page_zone(page);
6899 pfn = page_to_pfn(page);
6900 if (!zone_spans_pfn(zone, pfn))
6903 return !has_unmovable_pages(zone, page, 0, true);
6906 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6908 static unsigned long pfn_max_align_down(unsigned long pfn)
6910 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6911 pageblock_nr_pages) - 1);
6914 static unsigned long pfn_max_align_up(unsigned long pfn)
6916 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6917 pageblock_nr_pages));
6920 /* [start, end) must belong to a single zone. */
6921 static int __alloc_contig_migrate_range(struct compact_control *cc,
6922 unsigned long start, unsigned long end)
6924 /* This function is based on compact_zone() from compaction.c. */
6925 unsigned long nr_reclaimed;
6926 unsigned long pfn = start;
6927 unsigned int tries = 0;
6932 while (pfn < end || !list_empty(&cc->migratepages)) {
6933 if (fatal_signal_pending(current)) {
6938 if (list_empty(&cc->migratepages)) {
6939 cc->nr_migratepages = 0;
6940 pfn = isolate_migratepages_range(cc, pfn, end);
6946 } else if (++tries == 5) {
6947 ret = ret < 0 ? ret : -EBUSY;
6951 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6953 cc->nr_migratepages -= nr_reclaimed;
6955 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6956 NULL, 0, cc->mode, MR_CMA);
6959 putback_movable_pages(&cc->migratepages);
6966 * alloc_contig_range() -- tries to allocate given range of pages
6967 * @start: start PFN to allocate
6968 * @end: one-past-the-last PFN to allocate
6969 * @migratetype: migratetype of the underlaying pageblocks (either
6970 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6971 * in range must have the same migratetype and it must
6972 * be either of the two.
6974 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6975 * aligned, however it's the caller's responsibility to guarantee that
6976 * we are the only thread that changes migrate type of pageblocks the
6979 * The PFN range must belong to a single zone.
6981 * Returns zero on success or negative error code. On success all
6982 * pages which PFN is in [start, end) are allocated for the caller and
6983 * need to be freed with free_contig_range().
6985 int alloc_contig_range(unsigned long start, unsigned long end,
6986 unsigned migratetype)
6988 unsigned long outer_start, outer_end;
6992 struct compact_control cc = {
6993 .nr_migratepages = 0,
6995 .zone = page_zone(pfn_to_page(start)),
6996 .mode = MIGRATE_SYNC,
6997 .ignore_skip_hint = true,
6999 INIT_LIST_HEAD(&cc.migratepages);
7002 * What we do here is we mark all pageblocks in range as
7003 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7004 * have different sizes, and due to the way page allocator
7005 * work, we align the range to biggest of the two pages so
7006 * that page allocator won't try to merge buddies from
7007 * different pageblocks and change MIGRATE_ISOLATE to some
7008 * other migration type.
7010 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7011 * migrate the pages from an unaligned range (ie. pages that
7012 * we are interested in). This will put all the pages in
7013 * range back to page allocator as MIGRATE_ISOLATE.
7015 * When this is done, we take the pages in range from page
7016 * allocator removing them from the buddy system. This way
7017 * page allocator will never consider using them.
7019 * This lets us mark the pageblocks back as
7020 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7021 * aligned range but not in the unaligned, original range are
7022 * put back to page allocator so that buddy can use them.
7025 ret = start_isolate_page_range(pfn_max_align_down(start),
7026 pfn_max_align_up(end), migratetype,
7032 * In case of -EBUSY, we'd like to know which page causes problem.
7033 * So, just fall through. We will check it in test_pages_isolated().
7035 ret = __alloc_contig_migrate_range(&cc, start, end);
7036 if (ret && ret != -EBUSY)
7040 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7041 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7042 * more, all pages in [start, end) are free in page allocator.
7043 * What we are going to do is to allocate all pages from
7044 * [start, end) (that is remove them from page allocator).
7046 * The only problem is that pages at the beginning and at the
7047 * end of interesting range may be not aligned with pages that
7048 * page allocator holds, ie. they can be part of higher order
7049 * pages. Because of this, we reserve the bigger range and
7050 * once this is done free the pages we are not interested in.
7052 * We don't have to hold zone->lock here because the pages are
7053 * isolated thus they won't get removed from buddy.
7056 lru_add_drain_all();
7057 drain_all_pages(cc.zone);
7060 outer_start = start;
7061 while (!PageBuddy(pfn_to_page(outer_start))) {
7062 if (++order >= MAX_ORDER) {
7063 outer_start = start;
7066 outer_start &= ~0UL << order;
7069 if (outer_start != start) {
7070 order = page_order(pfn_to_page(outer_start));
7073 * outer_start page could be small order buddy page and
7074 * it doesn't include start page. Adjust outer_start
7075 * in this case to report failed page properly
7076 * on tracepoint in test_pages_isolated()
7078 if (outer_start + (1UL << order) <= start)
7079 outer_start = start;
7082 /* Make sure the range is really isolated. */
7083 if (test_pages_isolated(outer_start, end, false)) {
7084 pr_info("%s: [%lx, %lx) PFNs busy\n",
7085 __func__, outer_start, end);
7090 /* Grab isolated pages from freelists. */
7091 outer_end = isolate_freepages_range(&cc, outer_start, end);
7097 /* Free head and tail (if any) */
7098 if (start != outer_start)
7099 free_contig_range(outer_start, start - outer_start);
7100 if (end != outer_end)
7101 free_contig_range(end, outer_end - end);
7104 undo_isolate_page_range(pfn_max_align_down(start),
7105 pfn_max_align_up(end), migratetype);
7109 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7111 unsigned int count = 0;
7113 for (; nr_pages--; pfn++) {
7114 struct page *page = pfn_to_page(pfn);
7116 count += page_count(page) != 1;
7119 WARN(count != 0, "%d pages are still in use!\n", count);
7123 #ifdef CONFIG_MEMORY_HOTPLUG
7125 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7126 * page high values need to be recalulated.
7128 void __meminit zone_pcp_update(struct zone *zone)
7131 mutex_lock(&pcp_batch_high_lock);
7132 for_each_possible_cpu(cpu)
7133 pageset_set_high_and_batch(zone,
7134 per_cpu_ptr(zone->pageset, cpu));
7135 mutex_unlock(&pcp_batch_high_lock);
7139 void zone_pcp_reset(struct zone *zone)
7141 unsigned long flags;
7143 struct per_cpu_pageset *pset;
7145 /* avoid races with drain_pages() */
7146 local_irq_save(flags);
7147 if (zone->pageset != &boot_pageset) {
7148 for_each_online_cpu(cpu) {
7149 pset = per_cpu_ptr(zone->pageset, cpu);
7150 drain_zonestat(zone, pset);
7152 free_percpu(zone->pageset);
7153 zone->pageset = &boot_pageset;
7155 local_irq_restore(flags);
7158 #ifdef CONFIG_MEMORY_HOTREMOVE
7160 * All pages in the range must be isolated before calling this.
7163 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7167 unsigned int order, i;
7169 unsigned long flags;
7170 /* find the first valid pfn */
7171 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7176 zone = page_zone(pfn_to_page(pfn));
7177 spin_lock_irqsave(&zone->lock, flags);
7179 while (pfn < end_pfn) {
7180 if (!pfn_valid(pfn)) {
7184 page = pfn_to_page(pfn);
7186 * The HWPoisoned page may be not in buddy system, and
7187 * page_count() is not 0.
7189 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7191 SetPageReserved(page);
7195 BUG_ON(page_count(page));
7196 BUG_ON(!PageBuddy(page));
7197 order = page_order(page);
7198 #ifdef CONFIG_DEBUG_VM
7199 pr_info("remove from free list %lx %d %lx\n",
7200 pfn, 1 << order, end_pfn);
7202 list_del(&page->lru);
7203 rmv_page_order(page);
7204 zone->free_area[order].nr_free--;
7205 for (i = 0; i < (1 << order); i++)
7206 SetPageReserved((page+i));
7207 pfn += (1 << order);
7209 spin_unlock_irqrestore(&zone->lock, flags);
7213 bool is_free_buddy_page(struct page *page)
7215 struct zone *zone = page_zone(page);
7216 unsigned long pfn = page_to_pfn(page);
7217 unsigned long flags;
7220 spin_lock_irqsave(&zone->lock, flags);
7221 for (order = 0; order < MAX_ORDER; order++) {
7222 struct page *page_head = page - (pfn & ((1 << order) - 1));
7224 if (PageBuddy(page_head) && page_order(page_head) >= order)
7227 spin_unlock_irqrestore(&zone->lock, flags);
7229 return order < MAX_ORDER;