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/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_pages __read_mostly;
115 unsigned long totalcma_pages __read_mostly;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly;
124 int percpu_pagelist_fraction;
125 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex));
142 if (saved_gfp_mask) {
143 gfp_allowed_mask = saved_gfp_mask;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex));
151 WARN_ON(saved_gfp_mask);
152 saved_gfp_mask = gfp_allowed_mask;
153 gfp_allowed_mask &= ~GFP_IOFS;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly;
168 static void __free_pages_ok(struct page *page, unsigned int order);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages);
196 static char * const zone_names[MAX_NR_ZONES] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes = 1024;
211 int user_min_free_kbytes = -1;
213 static unsigned long __meminitdata nr_kernel_pages;
214 static unsigned long __meminitdata nr_all_pages;
215 static unsigned long __meminitdata dma_reserve;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __initdata required_kernelcore;
221 static unsigned long __initdata required_movablecore;
222 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly = MAX_NUMNODES;
231 int nr_online_nodes __read_mostly = 1;
232 EXPORT_SYMBOL(nr_node_ids);
233 EXPORT_SYMBOL(nr_online_nodes);
236 int page_group_by_mobility_disabled __read_mostly;
238 void set_pageblock_migratetype(struct page *page, int migratetype)
240 if (unlikely(page_group_by_mobility_disabled &&
241 migratetype < MIGRATE_PCPTYPES))
242 migratetype = MIGRATE_UNMOVABLE;
244 set_pageblock_flags_group(page, (unsigned long)migratetype,
245 PB_migrate, PB_migrate_end);
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn, zone_to_nid(zone), zone->name,
267 start_pfn, start_pfn + sp);
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page, const char *reason,
301 unsigned long bad_flags)
303 static unsigned long resume;
304 static unsigned long nr_shown;
305 static unsigned long nr_unshown;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page)) {
309 page_mapcount_reset(page); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown == 60) {
318 if (time_before(jiffies, resume)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume = jiffies + 60 * HZ;
333 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
334 current->comm, page_to_pfn(page));
335 dump_page_badflags(page, reason, bad_flags);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page *page)
362 __free_pages_ok(page, compound_order(page));
365 void prep_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
370 set_compound_page_dtor(page, free_compound_page);
371 set_compound_order(page, order);
373 for (i = 1; i < nr_pages; i++) {
374 struct page *p = page + i;
375 set_page_count(p, 0);
376 p->first_page = page;
377 /* Make sure p->first_page is always valid for PageTail() */
383 #ifdef CONFIG_DEBUG_PAGEALLOC
384 unsigned int _debug_guardpage_minorder;
385 bool _debug_pagealloc_enabled __read_mostly;
386 bool _debug_guardpage_enabled __read_mostly;
388 static int __init early_debug_pagealloc(char *buf)
393 if (strcmp(buf, "on") == 0)
394 _debug_pagealloc_enabled = true;
398 early_param("debug_pagealloc", early_debug_pagealloc);
400 static bool need_debug_guardpage(void)
402 /* If we don't use debug_pagealloc, we don't need guard page */
403 if (!debug_pagealloc_enabled())
409 static void init_debug_guardpage(void)
411 if (!debug_pagealloc_enabled())
414 _debug_guardpage_enabled = true;
417 struct page_ext_operations debug_guardpage_ops = {
418 .need = need_debug_guardpage,
419 .init = init_debug_guardpage,
422 static int __init debug_guardpage_minorder_setup(char *buf)
426 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
427 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
430 _debug_guardpage_minorder = res;
431 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
434 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
436 static inline void set_page_guard(struct zone *zone, struct page *page,
437 unsigned int order, int migratetype)
439 struct page_ext *page_ext;
441 if (!debug_guardpage_enabled())
444 page_ext = lookup_page_ext(page);
445 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
447 INIT_LIST_HEAD(&page->lru);
448 set_page_private(page, order);
449 /* Guard pages are not available for any usage */
450 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
453 static inline void clear_page_guard(struct zone *zone, struct page *page,
454 unsigned int order, int migratetype)
456 struct page_ext *page_ext;
458 if (!debug_guardpage_enabled())
461 page_ext = lookup_page_ext(page);
462 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
464 set_page_private(page, 0);
465 if (!is_migrate_isolate(migratetype))
466 __mod_zone_freepage_state(zone, (1 << order), migratetype);
469 struct page_ext_operations debug_guardpage_ops = { NULL, };
470 static inline void set_page_guard(struct zone *zone, struct page *page,
471 unsigned int order, int migratetype) {}
472 static inline void clear_page_guard(struct zone *zone, struct page *page,
473 unsigned int order, int migratetype) {}
476 static inline void set_page_order(struct page *page, unsigned int order)
478 set_page_private(page, order);
479 __SetPageBuddy(page);
482 static inline void rmv_page_order(struct page *page)
484 __ClearPageBuddy(page);
485 set_page_private(page, 0);
489 * This function checks whether a page is free && is the buddy
490 * we can do coalesce a page and its buddy if
491 * (a) the buddy is not in a hole &&
492 * (b) the buddy is in the buddy system &&
493 * (c) a page and its buddy have the same order &&
494 * (d) a page and its buddy are in the same zone.
496 * For recording whether a page is in the buddy system, we set ->_mapcount
497 * PAGE_BUDDY_MAPCOUNT_VALUE.
498 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
499 * serialized by zone->lock.
501 * For recording page's order, we use page_private(page).
503 static inline int page_is_buddy(struct page *page, struct page *buddy,
506 if (!pfn_valid_within(page_to_pfn(buddy)))
509 if (page_is_guard(buddy) && page_order(buddy) == order) {
510 if (page_zone_id(page) != page_zone_id(buddy))
513 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
518 if (PageBuddy(buddy) && page_order(buddy) == order) {
520 * zone check is done late to avoid uselessly
521 * calculating zone/node ids for pages that could
524 if (page_zone_id(page) != page_zone_id(buddy))
527 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
535 * Freeing function for a buddy system allocator.
537 * The concept of a buddy system is to maintain direct-mapped table
538 * (containing bit values) for memory blocks of various "orders".
539 * The bottom level table contains the map for the smallest allocatable
540 * units of memory (here, pages), and each level above it describes
541 * pairs of units from the levels below, hence, "buddies".
542 * At a high level, all that happens here is marking the table entry
543 * at the bottom level available, and propagating the changes upward
544 * as necessary, plus some accounting needed to play nicely with other
545 * parts of the VM system.
546 * At each level, we keep a list of pages, which are heads of continuous
547 * free pages of length of (1 << order) and marked with _mapcount
548 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
550 * So when we are allocating or freeing one, we can derive the state of the
551 * other. That is, if we allocate a small block, and both were
552 * free, the remainder of the region must be split into blocks.
553 * If a block is freed, and its buddy is also free, then this
554 * triggers coalescing into a block of larger size.
559 static inline void __free_one_page(struct page *page,
561 struct zone *zone, unsigned int order,
564 unsigned long page_idx;
565 unsigned long combined_idx;
566 unsigned long uninitialized_var(buddy_idx);
568 int max_order = MAX_ORDER;
570 VM_BUG_ON(!zone_is_initialized(zone));
571 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
573 VM_BUG_ON(migratetype == -1);
574 if (is_migrate_isolate(migratetype)) {
576 * We restrict max order of merging to prevent merge
577 * between freepages on isolate pageblock and normal
578 * pageblock. Without this, pageblock isolation
579 * could cause incorrect freepage accounting.
581 max_order = min(MAX_ORDER, pageblock_order + 1);
583 __mod_zone_freepage_state(zone, 1 << order, migratetype);
586 page_idx = pfn & ((1 << max_order) - 1);
588 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
589 VM_BUG_ON_PAGE(bad_range(zone, page), page);
591 while (order < max_order - 1) {
592 buddy_idx = __find_buddy_index(page_idx, order);
593 buddy = page + (buddy_idx - page_idx);
594 if (!page_is_buddy(page, buddy, order))
597 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
598 * merge with it and move up one order.
600 if (page_is_guard(buddy)) {
601 clear_page_guard(zone, buddy, order, migratetype);
603 list_del(&buddy->lru);
604 zone->free_area[order].nr_free--;
605 rmv_page_order(buddy);
607 combined_idx = buddy_idx & page_idx;
608 page = page + (combined_idx - page_idx);
609 page_idx = combined_idx;
612 set_page_order(page, order);
615 * If this is not the largest possible page, check if the buddy
616 * of the next-highest order is free. If it is, it's possible
617 * that pages are being freed that will coalesce soon. In case,
618 * that is happening, add the free page to the tail of the list
619 * so it's less likely to be used soon and more likely to be merged
620 * as a higher order page
622 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
623 struct page *higher_page, *higher_buddy;
624 combined_idx = buddy_idx & page_idx;
625 higher_page = page + (combined_idx - page_idx);
626 buddy_idx = __find_buddy_index(combined_idx, order + 1);
627 higher_buddy = higher_page + (buddy_idx - combined_idx);
628 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
629 list_add_tail(&page->lru,
630 &zone->free_area[order].free_list[migratetype]);
635 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
637 zone->free_area[order].nr_free++;
640 static inline int free_pages_check(struct page *page)
642 const char *bad_reason = NULL;
643 unsigned long bad_flags = 0;
645 if (unlikely(page_mapcount(page)))
646 bad_reason = "nonzero mapcount";
647 if (unlikely(page->mapping != NULL))
648 bad_reason = "non-NULL mapping";
649 if (unlikely(atomic_read(&page->_count) != 0))
650 bad_reason = "nonzero _count";
651 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
652 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
653 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
656 if (unlikely(page->mem_cgroup))
657 bad_reason = "page still charged to cgroup";
659 if (unlikely(bad_reason)) {
660 bad_page(page, bad_reason, bad_flags);
663 page_cpupid_reset_last(page);
664 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
665 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
670 * Frees a number of pages from the PCP lists
671 * Assumes all pages on list are in same zone, and of same order.
672 * count is the number of pages to free.
674 * If the zone was previously in an "all pages pinned" state then look to
675 * see if this freeing clears that state.
677 * And clear the zone's pages_scanned counter, to hold off the "all pages are
678 * pinned" detection logic.
680 static void free_pcppages_bulk(struct zone *zone, int count,
681 struct per_cpu_pages *pcp)
686 unsigned long nr_scanned;
688 spin_lock(&zone->lock);
689 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
691 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
695 struct list_head *list;
698 * Remove pages from lists in a round-robin fashion. A
699 * batch_free count is maintained that is incremented when an
700 * empty list is encountered. This is so more pages are freed
701 * off fuller lists instead of spinning excessively around empty
706 if (++migratetype == MIGRATE_PCPTYPES)
708 list = &pcp->lists[migratetype];
709 } while (list_empty(list));
711 /* This is the only non-empty list. Free them all. */
712 if (batch_free == MIGRATE_PCPTYPES)
713 batch_free = to_free;
716 int mt; /* migratetype of the to-be-freed page */
718 page = list_entry(list->prev, struct page, lru);
719 /* must delete as __free_one_page list manipulates */
720 list_del(&page->lru);
721 mt = get_freepage_migratetype(page);
722 if (unlikely(has_isolate_pageblock(zone)))
723 mt = get_pageblock_migratetype(page);
725 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
726 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
727 trace_mm_page_pcpu_drain(page, 0, mt);
728 } while (--to_free && --batch_free && !list_empty(list));
730 spin_unlock(&zone->lock);
733 static void free_one_page(struct zone *zone,
734 struct page *page, unsigned long pfn,
738 unsigned long nr_scanned;
739 spin_lock(&zone->lock);
740 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
742 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
744 if (unlikely(has_isolate_pageblock(zone) ||
745 is_migrate_isolate(migratetype))) {
746 migratetype = get_pfnblock_migratetype(page, pfn);
748 __free_one_page(page, pfn, zone, order, migratetype);
749 spin_unlock(&zone->lock);
752 static int free_tail_pages_check(struct page *head_page, struct page *page)
754 if (!IS_ENABLED(CONFIG_DEBUG_VM))
756 if (unlikely(!PageTail(page))) {
757 bad_page(page, "PageTail not set", 0);
760 if (unlikely(page->first_page != head_page)) {
761 bad_page(page, "first_page not consistent", 0);
767 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
768 unsigned long zone, int nid)
770 struct zone *z = &NODE_DATA(nid)->node_zones[zone];
772 set_page_links(page, zone, nid, pfn);
773 mminit_verify_page_links(page, zone, nid, pfn);
774 init_page_count(page);
775 page_mapcount_reset(page);
776 page_cpupid_reset_last(page);
779 * Mark the block movable so that blocks are reserved for
780 * movable at startup. This will force kernel allocations
781 * to reserve their blocks rather than leaking throughout
782 * the address space during boot when many long-lived
783 * kernel allocations are made. Later some blocks near
784 * the start are marked MIGRATE_RESERVE by
785 * setup_zone_migrate_reserve()
787 * bitmap is created for zone's valid pfn range. but memmap
788 * can be created for invalid pages (for alignment)
789 * check here not to call set_pageblock_migratetype() against
792 if ((z->zone_start_pfn <= pfn)
793 && (pfn < zone_end_pfn(z))
794 && !(pfn & (pageblock_nr_pages - 1)))
795 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
797 INIT_LIST_HEAD(&page->lru);
798 #ifdef WANT_PAGE_VIRTUAL
799 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
800 if (!is_highmem_idx(zone))
801 set_page_address(page, __va(pfn << PAGE_SHIFT));
805 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
808 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
812 * Initialised pages do not have PageReserved set. This function is
813 * called for each range allocated by the bootmem allocator and
814 * marks the pages PageReserved. The remaining valid pages are later
815 * sent to the buddy page allocator.
817 void reserve_bootmem_region(unsigned long start, unsigned long end)
819 unsigned long start_pfn = PFN_DOWN(start);
820 unsigned long end_pfn = PFN_UP(end);
822 for (; start_pfn < end_pfn; start_pfn++)
823 if (pfn_valid(start_pfn))
824 SetPageReserved(pfn_to_page(start_pfn));
827 static bool free_pages_prepare(struct page *page, unsigned int order)
829 bool compound = PageCompound(page);
832 VM_BUG_ON_PAGE(PageTail(page), page);
833 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
835 trace_mm_page_free(page, order);
836 kmemcheck_free_shadow(page, order);
837 kasan_free_pages(page, order);
840 page->mapping = NULL;
841 bad += free_pages_check(page);
842 for (i = 1; i < (1 << order); i++) {
844 bad += free_tail_pages_check(page, page + i);
845 bad += free_pages_check(page + i);
850 reset_page_owner(page, order);
852 if (!PageHighMem(page)) {
853 debug_check_no_locks_freed(page_address(page),
855 debug_check_no_obj_freed(page_address(page),
858 arch_free_page(page, order);
859 kernel_map_pages(page, 1 << order, 0);
864 static void __free_pages_ok(struct page *page, unsigned int order)
868 unsigned long pfn = page_to_pfn(page);
870 if (!free_pages_prepare(page, order))
873 migratetype = get_pfnblock_migratetype(page, pfn);
874 local_irq_save(flags);
875 __count_vm_events(PGFREE, 1 << order);
876 set_freepage_migratetype(page, migratetype);
877 free_one_page(page_zone(page), page, pfn, order, migratetype);
878 local_irq_restore(flags);
881 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
884 unsigned int nr_pages = 1 << order;
885 struct page *p = page;
889 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
891 __ClearPageReserved(p);
892 set_page_count(p, 0);
894 __ClearPageReserved(p);
895 set_page_count(p, 0);
897 page_zone(page)->managed_pages += nr_pages;
898 set_page_refcounted(page);
899 __free_pages(page, order);
902 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
903 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
904 /* Only safe to use early in boot when initialisation is single-threaded */
905 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
907 int __meminit early_pfn_to_nid(unsigned long pfn)
911 /* The system will behave unpredictably otherwise */
912 BUG_ON(system_state != SYSTEM_BOOTING);
914 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
922 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
923 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
924 struct mminit_pfnnid_cache *state)
928 nid = __early_pfn_to_nid(pfn, state);
929 if (nid >= 0 && nid != node)
934 /* Only safe to use early in boot when initialisation is single-threaded */
935 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
937 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
942 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
946 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
947 struct mminit_pfnnid_cache *state)
955 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
956 void __init init_cma_reserved_pageblock(struct page *page)
958 unsigned i = pageblock_nr_pages;
959 struct page *p = page;
962 __ClearPageReserved(p);
963 set_page_count(p, 0);
966 set_pageblock_migratetype(page, MIGRATE_CMA);
968 if (pageblock_order >= MAX_ORDER) {
969 i = pageblock_nr_pages;
972 set_page_refcounted(p);
973 __free_pages(p, MAX_ORDER - 1);
974 p += MAX_ORDER_NR_PAGES;
975 } while (i -= MAX_ORDER_NR_PAGES);
977 set_page_refcounted(page);
978 __free_pages(page, pageblock_order);
981 adjust_managed_page_count(page, pageblock_nr_pages);
986 * The order of subdivision here is critical for the IO subsystem.
987 * Please do not alter this order without good reasons and regression
988 * testing. Specifically, as large blocks of memory are subdivided,
989 * the order in which smaller blocks are delivered depends on the order
990 * they're subdivided in this function. This is the primary factor
991 * influencing the order in which pages are delivered to the IO
992 * subsystem according to empirical testing, and this is also justified
993 * by considering the behavior of a buddy system containing a single
994 * large block of memory acted on by a series of small allocations.
995 * This behavior is a critical factor in sglist merging's success.
999 static inline void expand(struct zone *zone, struct page *page,
1000 int low, int high, struct free_area *area,
1003 unsigned long size = 1 << high;
1005 while (high > low) {
1009 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1011 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1012 debug_guardpage_enabled() &&
1013 high < debug_guardpage_minorder()) {
1015 * Mark as guard pages (or page), that will allow to
1016 * merge back to allocator when buddy will be freed.
1017 * Corresponding page table entries will not be touched,
1018 * pages will stay not present in virtual address space
1020 set_page_guard(zone, &page[size], high, migratetype);
1023 list_add(&page[size].lru, &area->free_list[migratetype]);
1025 set_page_order(&page[size], high);
1030 * This page is about to be returned from the page allocator
1032 static inline int check_new_page(struct page *page)
1034 const char *bad_reason = NULL;
1035 unsigned long bad_flags = 0;
1037 if (unlikely(page_mapcount(page)))
1038 bad_reason = "nonzero mapcount";
1039 if (unlikely(page->mapping != NULL))
1040 bad_reason = "non-NULL mapping";
1041 if (unlikely(atomic_read(&page->_count) != 0))
1042 bad_reason = "nonzero _count";
1043 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1044 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1045 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1048 if (unlikely(page->mem_cgroup))
1049 bad_reason = "page still charged to cgroup";
1051 if (unlikely(bad_reason)) {
1052 bad_page(page, bad_reason, bad_flags);
1058 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1063 for (i = 0; i < (1 << order); i++) {
1064 struct page *p = page + i;
1065 if (unlikely(check_new_page(p)))
1069 set_page_private(page, 0);
1070 set_page_refcounted(page);
1072 arch_alloc_page(page, order);
1073 kernel_map_pages(page, 1 << order, 1);
1074 kasan_alloc_pages(page, order);
1076 if (gfp_flags & __GFP_ZERO)
1077 for (i = 0; i < (1 << order); i++)
1078 clear_highpage(page + i);
1080 if (order && (gfp_flags & __GFP_COMP))
1081 prep_compound_page(page, order);
1083 set_page_owner(page, order, gfp_flags);
1086 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1087 * allocate the page. The expectation is that the caller is taking
1088 * steps that will free more memory. The caller should avoid the page
1089 * being used for !PFMEMALLOC purposes.
1091 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1097 * Go through the free lists for the given migratetype and remove
1098 * the smallest available page from the freelists
1101 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1104 unsigned int current_order;
1105 struct free_area *area;
1108 /* Find a page of the appropriate size in the preferred list */
1109 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1110 area = &(zone->free_area[current_order]);
1111 if (list_empty(&area->free_list[migratetype]))
1114 page = list_entry(area->free_list[migratetype].next,
1116 list_del(&page->lru);
1117 rmv_page_order(page);
1119 expand(zone, page, order, current_order, area, migratetype);
1120 set_freepage_migratetype(page, migratetype);
1129 * This array describes the order lists are fallen back to when
1130 * the free lists for the desirable migrate type are depleted
1132 static int fallbacks[MIGRATE_TYPES][4] = {
1133 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1134 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1135 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1137 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1139 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1140 #ifdef CONFIG_MEMORY_ISOLATION
1141 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1146 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1149 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1152 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1153 unsigned int order) { return NULL; }
1157 * Move the free pages in a range to the free lists of the requested type.
1158 * Note that start_page and end_pages are not aligned on a pageblock
1159 * boundary. If alignment is required, use move_freepages_block()
1161 int move_freepages(struct zone *zone,
1162 struct page *start_page, struct page *end_page,
1166 unsigned long order;
1167 int pages_moved = 0;
1169 #ifndef CONFIG_HOLES_IN_ZONE
1171 * page_zone is not safe to call in this context when
1172 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1173 * anyway as we check zone boundaries in move_freepages_block().
1174 * Remove at a later date when no bug reports exist related to
1175 * grouping pages by mobility
1177 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1180 for (page = start_page; page <= end_page;) {
1181 /* Make sure we are not inadvertently changing nodes */
1182 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1184 if (!pfn_valid_within(page_to_pfn(page))) {
1189 if (!PageBuddy(page)) {
1194 order = page_order(page);
1195 list_move(&page->lru,
1196 &zone->free_area[order].free_list[migratetype]);
1197 set_freepage_migratetype(page, migratetype);
1199 pages_moved += 1 << order;
1205 int move_freepages_block(struct zone *zone, struct page *page,
1208 unsigned long start_pfn, end_pfn;
1209 struct page *start_page, *end_page;
1211 start_pfn = page_to_pfn(page);
1212 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1213 start_page = pfn_to_page(start_pfn);
1214 end_page = start_page + pageblock_nr_pages - 1;
1215 end_pfn = start_pfn + pageblock_nr_pages - 1;
1217 /* Do not cross zone boundaries */
1218 if (!zone_spans_pfn(zone, start_pfn))
1220 if (!zone_spans_pfn(zone, end_pfn))
1223 return move_freepages(zone, start_page, end_page, migratetype);
1226 static void change_pageblock_range(struct page *pageblock_page,
1227 int start_order, int migratetype)
1229 int nr_pageblocks = 1 << (start_order - pageblock_order);
1231 while (nr_pageblocks--) {
1232 set_pageblock_migratetype(pageblock_page, migratetype);
1233 pageblock_page += pageblock_nr_pages;
1238 * When we are falling back to another migratetype during allocation, try to
1239 * steal extra free pages from the same pageblocks to satisfy further
1240 * allocations, instead of polluting multiple pageblocks.
1242 * If we are stealing a relatively large buddy page, it is likely there will
1243 * be more free pages in the pageblock, so try to steal them all. For
1244 * reclaimable and unmovable allocations, we steal regardless of page size,
1245 * as fragmentation caused by those allocations polluting movable pageblocks
1246 * is worse than movable allocations stealing from unmovable and reclaimable
1249 static bool can_steal_fallback(unsigned int order, int start_mt)
1252 * Leaving this order check is intended, although there is
1253 * relaxed order check in next check. The reason is that
1254 * we can actually steal whole pageblock if this condition met,
1255 * but, below check doesn't guarantee it and that is just heuristic
1256 * so could be changed anytime.
1258 if (order >= pageblock_order)
1261 if (order >= pageblock_order / 2 ||
1262 start_mt == MIGRATE_RECLAIMABLE ||
1263 start_mt == MIGRATE_UNMOVABLE ||
1264 page_group_by_mobility_disabled)
1271 * This function implements actual steal behaviour. If order is large enough,
1272 * we can steal whole pageblock. If not, we first move freepages in this
1273 * pageblock and check whether half of pages are moved or not. If half of
1274 * pages are moved, we can change migratetype of pageblock and permanently
1275 * use it's pages as requested migratetype in the future.
1277 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1280 int current_order = page_order(page);
1283 /* Take ownership for orders >= pageblock_order */
1284 if (current_order >= pageblock_order) {
1285 change_pageblock_range(page, current_order, start_type);
1289 pages = move_freepages_block(zone, page, start_type);
1291 /* Claim the whole block if over half of it is free */
1292 if (pages >= (1 << (pageblock_order-1)) ||
1293 page_group_by_mobility_disabled)
1294 set_pageblock_migratetype(page, start_type);
1298 * Check whether there is a suitable fallback freepage with requested order.
1299 * If only_stealable is true, this function returns fallback_mt only if
1300 * we can steal other freepages all together. This would help to reduce
1301 * fragmentation due to mixed migratetype pages in one pageblock.
1303 int find_suitable_fallback(struct free_area *area, unsigned int order,
1304 int migratetype, bool only_stealable, bool *can_steal)
1309 if (area->nr_free == 0)
1314 fallback_mt = fallbacks[migratetype][i];
1315 if (fallback_mt == MIGRATE_RESERVE)
1318 if (list_empty(&area->free_list[fallback_mt]))
1321 if (can_steal_fallback(order, migratetype))
1324 if (!only_stealable)
1334 /* Remove an element from the buddy allocator from the fallback list */
1335 static inline struct page *
1336 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1338 struct free_area *area;
1339 unsigned int current_order;
1344 /* Find the largest possible block of pages in the other list */
1345 for (current_order = MAX_ORDER-1;
1346 current_order >= order && current_order <= MAX_ORDER-1;
1348 area = &(zone->free_area[current_order]);
1349 fallback_mt = find_suitable_fallback(area, current_order,
1350 start_migratetype, false, &can_steal);
1351 if (fallback_mt == -1)
1354 page = list_entry(area->free_list[fallback_mt].next,
1357 steal_suitable_fallback(zone, page, start_migratetype);
1359 /* Remove the page from the freelists */
1361 list_del(&page->lru);
1362 rmv_page_order(page);
1364 expand(zone, page, order, current_order, area,
1367 * The freepage_migratetype may differ from pageblock's
1368 * migratetype depending on the decisions in
1369 * try_to_steal_freepages(). This is OK as long as it
1370 * does not differ for MIGRATE_CMA pageblocks. For CMA
1371 * we need to make sure unallocated pages flushed from
1372 * pcp lists are returned to the correct freelist.
1374 set_freepage_migratetype(page, start_migratetype);
1376 trace_mm_page_alloc_extfrag(page, order, current_order,
1377 start_migratetype, fallback_mt);
1386 * Do the hard work of removing an element from the buddy allocator.
1387 * Call me with the zone->lock already held.
1389 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1395 page = __rmqueue_smallest(zone, order, migratetype);
1397 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1398 if (migratetype == MIGRATE_MOVABLE)
1399 page = __rmqueue_cma_fallback(zone, order);
1402 page = __rmqueue_fallback(zone, order, migratetype);
1405 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1406 * is used because __rmqueue_smallest is an inline function
1407 * and we want just one call site
1410 migratetype = MIGRATE_RESERVE;
1415 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1420 * Obtain a specified number of elements from the buddy allocator, all under
1421 * a single hold of the lock, for efficiency. Add them to the supplied list.
1422 * Returns the number of new pages which were placed at *list.
1424 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1425 unsigned long count, struct list_head *list,
1426 int migratetype, bool cold)
1430 spin_lock(&zone->lock);
1431 for (i = 0; i < count; ++i) {
1432 struct page *page = __rmqueue(zone, order, migratetype);
1433 if (unlikely(page == NULL))
1437 * Split buddy pages returned by expand() are received here
1438 * in physical page order. The page is added to the callers and
1439 * list and the list head then moves forward. From the callers
1440 * perspective, the linked list is ordered by page number in
1441 * some conditions. This is useful for IO devices that can
1442 * merge IO requests if the physical pages are ordered
1446 list_add(&page->lru, list);
1448 list_add_tail(&page->lru, list);
1450 if (is_migrate_cma(get_freepage_migratetype(page)))
1451 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1454 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1455 spin_unlock(&zone->lock);
1461 * Called from the vmstat counter updater to drain pagesets of this
1462 * currently executing processor on remote nodes after they have
1465 * Note that this function must be called with the thread pinned to
1466 * a single processor.
1468 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1470 unsigned long flags;
1471 int to_drain, batch;
1473 local_irq_save(flags);
1474 batch = READ_ONCE(pcp->batch);
1475 to_drain = min(pcp->count, batch);
1477 free_pcppages_bulk(zone, to_drain, pcp);
1478 pcp->count -= to_drain;
1480 local_irq_restore(flags);
1485 * Drain pcplists of the indicated processor and zone.
1487 * The processor must either be the current processor and the
1488 * thread pinned to the current processor or a processor that
1491 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1493 unsigned long flags;
1494 struct per_cpu_pageset *pset;
1495 struct per_cpu_pages *pcp;
1497 local_irq_save(flags);
1498 pset = per_cpu_ptr(zone->pageset, cpu);
1502 free_pcppages_bulk(zone, pcp->count, pcp);
1505 local_irq_restore(flags);
1509 * Drain pcplists of all zones on the indicated processor.
1511 * The processor must either be the current processor and the
1512 * thread pinned to the current processor or a processor that
1515 static void drain_pages(unsigned int cpu)
1519 for_each_populated_zone(zone) {
1520 drain_pages_zone(cpu, zone);
1525 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1527 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1528 * the single zone's pages.
1530 void drain_local_pages(struct zone *zone)
1532 int cpu = smp_processor_id();
1535 drain_pages_zone(cpu, zone);
1541 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1543 * When zone parameter is non-NULL, spill just the single zone's pages.
1545 * Note that this code is protected against sending an IPI to an offline
1546 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1547 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1548 * nothing keeps CPUs from showing up after we populated the cpumask and
1549 * before the call to on_each_cpu_mask().
1551 void drain_all_pages(struct zone *zone)
1556 * Allocate in the BSS so we wont require allocation in
1557 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1559 static cpumask_t cpus_with_pcps;
1562 * We don't care about racing with CPU hotplug event
1563 * as offline notification will cause the notified
1564 * cpu to drain that CPU pcps and on_each_cpu_mask
1565 * disables preemption as part of its processing
1567 for_each_online_cpu(cpu) {
1568 struct per_cpu_pageset *pcp;
1570 bool has_pcps = false;
1573 pcp = per_cpu_ptr(zone->pageset, cpu);
1577 for_each_populated_zone(z) {
1578 pcp = per_cpu_ptr(z->pageset, cpu);
1579 if (pcp->pcp.count) {
1587 cpumask_set_cpu(cpu, &cpus_with_pcps);
1589 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1591 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1595 #ifdef CONFIG_HIBERNATION
1597 void mark_free_pages(struct zone *zone)
1599 unsigned long pfn, max_zone_pfn;
1600 unsigned long flags;
1601 unsigned int order, t;
1602 struct list_head *curr;
1604 if (zone_is_empty(zone))
1607 spin_lock_irqsave(&zone->lock, flags);
1609 max_zone_pfn = zone_end_pfn(zone);
1610 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1611 if (pfn_valid(pfn)) {
1612 struct page *page = pfn_to_page(pfn);
1614 if (!swsusp_page_is_forbidden(page))
1615 swsusp_unset_page_free(page);
1618 for_each_migratetype_order(order, t) {
1619 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1622 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1623 for (i = 0; i < (1UL << order); i++)
1624 swsusp_set_page_free(pfn_to_page(pfn + i));
1627 spin_unlock_irqrestore(&zone->lock, flags);
1629 #endif /* CONFIG_PM */
1632 * Free a 0-order page
1633 * cold == true ? free a cold page : free a hot page
1635 void free_hot_cold_page(struct page *page, bool cold)
1637 struct zone *zone = page_zone(page);
1638 struct per_cpu_pages *pcp;
1639 unsigned long flags;
1640 unsigned long pfn = page_to_pfn(page);
1643 if (!free_pages_prepare(page, 0))
1646 migratetype = get_pfnblock_migratetype(page, pfn);
1647 set_freepage_migratetype(page, migratetype);
1648 local_irq_save(flags);
1649 __count_vm_event(PGFREE);
1652 * We only track unmovable, reclaimable and movable on pcp lists.
1653 * Free ISOLATE pages back to the allocator because they are being
1654 * offlined but treat RESERVE as movable pages so we can get those
1655 * areas back if necessary. Otherwise, we may have to free
1656 * excessively into the page allocator
1658 if (migratetype >= MIGRATE_PCPTYPES) {
1659 if (unlikely(is_migrate_isolate(migratetype))) {
1660 free_one_page(zone, page, pfn, 0, migratetype);
1663 migratetype = MIGRATE_MOVABLE;
1666 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1668 list_add(&page->lru, &pcp->lists[migratetype]);
1670 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1672 if (pcp->count >= pcp->high) {
1673 unsigned long batch = READ_ONCE(pcp->batch);
1674 free_pcppages_bulk(zone, batch, pcp);
1675 pcp->count -= batch;
1679 local_irq_restore(flags);
1683 * Free a list of 0-order pages
1685 void free_hot_cold_page_list(struct list_head *list, bool cold)
1687 struct page *page, *next;
1689 list_for_each_entry_safe(page, next, list, lru) {
1690 trace_mm_page_free_batched(page, cold);
1691 free_hot_cold_page(page, cold);
1696 * split_page takes a non-compound higher-order page, and splits it into
1697 * n (1<<order) sub-pages: page[0..n]
1698 * Each sub-page must be freed individually.
1700 * Note: this is probably too low level an operation for use in drivers.
1701 * Please consult with lkml before using this in your driver.
1703 void split_page(struct page *page, unsigned int order)
1707 VM_BUG_ON_PAGE(PageCompound(page), page);
1708 VM_BUG_ON_PAGE(!page_count(page), page);
1710 #ifdef CONFIG_KMEMCHECK
1712 * Split shadow pages too, because free(page[0]) would
1713 * otherwise free the whole shadow.
1715 if (kmemcheck_page_is_tracked(page))
1716 split_page(virt_to_page(page[0].shadow), order);
1719 set_page_owner(page, 0, 0);
1720 for (i = 1; i < (1 << order); i++) {
1721 set_page_refcounted(page + i);
1722 set_page_owner(page + i, 0, 0);
1725 EXPORT_SYMBOL_GPL(split_page);
1727 int __isolate_free_page(struct page *page, unsigned int order)
1729 unsigned long watermark;
1733 BUG_ON(!PageBuddy(page));
1735 zone = page_zone(page);
1736 mt = get_pageblock_migratetype(page);
1738 if (!is_migrate_isolate(mt)) {
1739 /* Obey watermarks as if the page was being allocated */
1740 watermark = low_wmark_pages(zone) + (1 << order);
1741 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1744 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1747 /* Remove page from free list */
1748 list_del(&page->lru);
1749 zone->free_area[order].nr_free--;
1750 rmv_page_order(page);
1752 /* Set the pageblock if the isolated page is at least a pageblock */
1753 if (order >= pageblock_order - 1) {
1754 struct page *endpage = page + (1 << order) - 1;
1755 for (; page < endpage; page += pageblock_nr_pages) {
1756 int mt = get_pageblock_migratetype(page);
1757 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1758 set_pageblock_migratetype(page,
1763 set_page_owner(page, order, 0);
1764 return 1UL << order;
1768 * Similar to split_page except the page is already free. As this is only
1769 * being used for migration, the migratetype of the block also changes.
1770 * As this is called with interrupts disabled, the caller is responsible
1771 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1774 * Note: this is probably too low level an operation for use in drivers.
1775 * Please consult with lkml before using this in your driver.
1777 int split_free_page(struct page *page)
1782 order = page_order(page);
1784 nr_pages = __isolate_free_page(page, order);
1788 /* Split into individual pages */
1789 set_page_refcounted(page);
1790 split_page(page, order);
1795 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1798 struct page *buffered_rmqueue(struct zone *preferred_zone,
1799 struct zone *zone, unsigned int order,
1800 gfp_t gfp_flags, int migratetype)
1802 unsigned long flags;
1804 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1806 if (likely(order == 0)) {
1807 struct per_cpu_pages *pcp;
1808 struct list_head *list;
1810 local_irq_save(flags);
1811 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1812 list = &pcp->lists[migratetype];
1813 if (list_empty(list)) {
1814 pcp->count += rmqueue_bulk(zone, 0,
1817 if (unlikely(list_empty(list)))
1822 page = list_entry(list->prev, struct page, lru);
1824 page = list_entry(list->next, struct page, lru);
1826 list_del(&page->lru);
1829 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1831 * __GFP_NOFAIL is not to be used in new code.
1833 * All __GFP_NOFAIL callers should be fixed so that they
1834 * properly detect and handle allocation failures.
1836 * We most definitely don't want callers attempting to
1837 * allocate greater than order-1 page units with
1840 WARN_ON_ONCE(order > 1);
1842 spin_lock_irqsave(&zone->lock, flags);
1843 page = __rmqueue(zone, order, migratetype);
1844 spin_unlock(&zone->lock);
1847 __mod_zone_freepage_state(zone, -(1 << order),
1848 get_freepage_migratetype(page));
1851 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1852 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1853 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1854 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1856 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1857 zone_statistics(preferred_zone, zone, gfp_flags);
1858 local_irq_restore(flags);
1860 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1864 local_irq_restore(flags);
1868 #ifdef CONFIG_FAIL_PAGE_ALLOC
1871 struct fault_attr attr;
1873 u32 ignore_gfp_highmem;
1874 u32 ignore_gfp_wait;
1876 } fail_page_alloc = {
1877 .attr = FAULT_ATTR_INITIALIZER,
1878 .ignore_gfp_wait = 1,
1879 .ignore_gfp_highmem = 1,
1883 static int __init setup_fail_page_alloc(char *str)
1885 return setup_fault_attr(&fail_page_alloc.attr, str);
1887 __setup("fail_page_alloc=", setup_fail_page_alloc);
1889 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1891 if (order < fail_page_alloc.min_order)
1893 if (gfp_mask & __GFP_NOFAIL)
1895 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1897 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1900 return should_fail(&fail_page_alloc.attr, 1 << order);
1903 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1905 static int __init fail_page_alloc_debugfs(void)
1907 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1910 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1911 &fail_page_alloc.attr);
1913 return PTR_ERR(dir);
1915 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1916 &fail_page_alloc.ignore_gfp_wait))
1918 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1919 &fail_page_alloc.ignore_gfp_highmem))
1921 if (!debugfs_create_u32("min-order", mode, dir,
1922 &fail_page_alloc.min_order))
1927 debugfs_remove_recursive(dir);
1932 late_initcall(fail_page_alloc_debugfs);
1934 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1936 #else /* CONFIG_FAIL_PAGE_ALLOC */
1938 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1943 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1946 * Return true if free pages are above 'mark'. This takes into account the order
1947 * of the allocation.
1949 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1950 unsigned long mark, int classzone_idx, int alloc_flags,
1953 /* free_pages may go negative - that's OK */
1958 free_pages -= (1 << order) - 1;
1959 if (alloc_flags & ALLOC_HIGH)
1961 if (alloc_flags & ALLOC_HARDER)
1964 /* If allocation can't use CMA areas don't use free CMA pages */
1965 if (!(alloc_flags & ALLOC_CMA))
1966 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1969 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1971 for (o = 0; o < order; o++) {
1972 /* At the next order, this order's pages become unavailable */
1973 free_pages -= z->free_area[o].nr_free << o;
1975 /* Require fewer higher order pages to be free */
1978 if (free_pages <= min)
1984 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1985 int classzone_idx, int alloc_flags)
1987 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1988 zone_page_state(z, NR_FREE_PAGES));
1991 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1992 unsigned long mark, int classzone_idx, int alloc_flags)
1994 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1996 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1997 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1999 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2005 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2006 * skip over zones that are not allowed by the cpuset, or that have
2007 * been recently (in last second) found to be nearly full. See further
2008 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2009 * that have to skip over a lot of full or unallowed zones.
2011 * If the zonelist cache is present in the passed zonelist, then
2012 * returns a pointer to the allowed node mask (either the current
2013 * tasks mems_allowed, or node_states[N_MEMORY].)
2015 * If the zonelist cache is not available for this zonelist, does
2016 * nothing and returns NULL.
2018 * If the fullzones BITMAP in the zonelist cache is stale (more than
2019 * a second since last zap'd) then we zap it out (clear its bits.)
2021 * We hold off even calling zlc_setup, until after we've checked the
2022 * first zone in the zonelist, on the theory that most allocations will
2023 * be satisfied from that first zone, so best to examine that zone as
2024 * quickly as we can.
2026 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2028 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2029 nodemask_t *allowednodes; /* zonelist_cache approximation */
2031 zlc = zonelist->zlcache_ptr;
2035 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2036 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2037 zlc->last_full_zap = jiffies;
2040 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2041 &cpuset_current_mems_allowed :
2042 &node_states[N_MEMORY];
2043 return allowednodes;
2047 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2048 * if it is worth looking at further for free memory:
2049 * 1) Check that the zone isn't thought to be full (doesn't have its
2050 * bit set in the zonelist_cache fullzones BITMAP).
2051 * 2) Check that the zones node (obtained from the zonelist_cache
2052 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2053 * Return true (non-zero) if zone is worth looking at further, or
2054 * else return false (zero) if it is not.
2056 * This check -ignores- the distinction between various watermarks,
2057 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2058 * found to be full for any variation of these watermarks, it will
2059 * be considered full for up to one second by all requests, unless
2060 * we are so low on memory on all allowed nodes that we are forced
2061 * into the second scan of the zonelist.
2063 * In the second scan we ignore this zonelist cache and exactly
2064 * apply the watermarks to all zones, even it is slower to do so.
2065 * We are low on memory in the second scan, and should leave no stone
2066 * unturned looking for a free page.
2068 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2069 nodemask_t *allowednodes)
2071 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2072 int i; /* index of *z in zonelist zones */
2073 int n; /* node that zone *z is on */
2075 zlc = zonelist->zlcache_ptr;
2079 i = z - zonelist->_zonerefs;
2082 /* This zone is worth trying if it is allowed but not full */
2083 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2087 * Given 'z' scanning a zonelist, set the corresponding bit in
2088 * zlc->fullzones, so that subsequent attempts to allocate a page
2089 * from that zone don't waste time re-examining it.
2091 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2093 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2094 int i; /* index of *z in zonelist zones */
2096 zlc = zonelist->zlcache_ptr;
2100 i = z - zonelist->_zonerefs;
2102 set_bit(i, zlc->fullzones);
2106 * clear all zones full, called after direct reclaim makes progress so that
2107 * a zone that was recently full is not skipped over for up to a second
2109 static void zlc_clear_zones_full(struct zonelist *zonelist)
2111 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2113 zlc = zonelist->zlcache_ptr;
2117 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2120 static bool zone_local(struct zone *local_zone, struct zone *zone)
2122 return local_zone->node == zone->node;
2125 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2127 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2131 #else /* CONFIG_NUMA */
2133 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2138 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2139 nodemask_t *allowednodes)
2144 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2148 static void zlc_clear_zones_full(struct zonelist *zonelist)
2152 static bool zone_local(struct zone *local_zone, struct zone *zone)
2157 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2162 #endif /* CONFIG_NUMA */
2164 static void reset_alloc_batches(struct zone *preferred_zone)
2166 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2169 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2170 high_wmark_pages(zone) - low_wmark_pages(zone) -
2171 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2172 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2173 } while (zone++ != preferred_zone);
2177 * get_page_from_freelist goes through the zonelist trying to allocate
2180 static struct page *
2181 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2182 const struct alloc_context *ac)
2184 struct zonelist *zonelist = ac->zonelist;
2186 struct page *page = NULL;
2188 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2189 int zlc_active = 0; /* set if using zonelist_cache */
2190 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2191 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2192 (gfp_mask & __GFP_WRITE);
2193 int nr_fair_skipped = 0;
2194 bool zonelist_rescan;
2197 zonelist_rescan = false;
2200 * Scan zonelist, looking for a zone with enough free.
2201 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2203 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2207 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2208 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2210 if (cpusets_enabled() &&
2211 (alloc_flags & ALLOC_CPUSET) &&
2212 !cpuset_zone_allowed(zone, gfp_mask))
2215 * Distribute pages in proportion to the individual
2216 * zone size to ensure fair page aging. The zone a
2217 * page was allocated in should have no effect on the
2218 * time the page has in memory before being reclaimed.
2220 if (alloc_flags & ALLOC_FAIR) {
2221 if (!zone_local(ac->preferred_zone, zone))
2223 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2229 * When allocating a page cache page for writing, we
2230 * want to get it from a zone that is within its dirty
2231 * limit, such that no single zone holds more than its
2232 * proportional share of globally allowed dirty pages.
2233 * The dirty limits take into account the zone's
2234 * lowmem reserves and high watermark so that kswapd
2235 * should be able to balance it without having to
2236 * write pages from its LRU list.
2238 * This may look like it could increase pressure on
2239 * lower zones by failing allocations in higher zones
2240 * before they are full. But the pages that do spill
2241 * over are limited as the lower zones are protected
2242 * by this very same mechanism. It should not become
2243 * a practical burden to them.
2245 * XXX: For now, allow allocations to potentially
2246 * exceed the per-zone dirty limit in the slowpath
2247 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2248 * which is important when on a NUMA setup the allowed
2249 * zones are together not big enough to reach the
2250 * global limit. The proper fix for these situations
2251 * will require awareness of zones in the
2252 * dirty-throttling and the flusher threads.
2254 if (consider_zone_dirty && !zone_dirty_ok(zone))
2257 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2258 if (!zone_watermark_ok(zone, order, mark,
2259 ac->classzone_idx, alloc_flags)) {
2262 /* Checked here to keep the fast path fast */
2263 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2264 if (alloc_flags & ALLOC_NO_WATERMARKS)
2267 if (IS_ENABLED(CONFIG_NUMA) &&
2268 !did_zlc_setup && nr_online_nodes > 1) {
2270 * we do zlc_setup if there are multiple nodes
2271 * and before considering the first zone allowed
2274 allowednodes = zlc_setup(zonelist, alloc_flags);
2279 if (zone_reclaim_mode == 0 ||
2280 !zone_allows_reclaim(ac->preferred_zone, zone))
2281 goto this_zone_full;
2284 * As we may have just activated ZLC, check if the first
2285 * eligible zone has failed zone_reclaim recently.
2287 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2288 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2291 ret = zone_reclaim(zone, gfp_mask, order);
2293 case ZONE_RECLAIM_NOSCAN:
2296 case ZONE_RECLAIM_FULL:
2297 /* scanned but unreclaimable */
2300 /* did we reclaim enough */
2301 if (zone_watermark_ok(zone, order, mark,
2302 ac->classzone_idx, alloc_flags))
2306 * Failed to reclaim enough to meet watermark.
2307 * Only mark the zone full if checking the min
2308 * watermark or if we failed to reclaim just
2309 * 1<<order pages or else the page allocator
2310 * fastpath will prematurely mark zones full
2311 * when the watermark is between the low and
2314 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2315 ret == ZONE_RECLAIM_SOME)
2316 goto this_zone_full;
2323 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2324 gfp_mask, ac->migratetype);
2326 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2331 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2332 zlc_mark_zone_full(zonelist, z);
2336 * The first pass makes sure allocations are spread fairly within the
2337 * local node. However, the local node might have free pages left
2338 * after the fairness batches are exhausted, and remote zones haven't
2339 * even been considered yet. Try once more without fairness, and
2340 * include remote zones now, before entering the slowpath and waking
2341 * kswapd: prefer spilling to a remote zone over swapping locally.
2343 if (alloc_flags & ALLOC_FAIR) {
2344 alloc_flags &= ~ALLOC_FAIR;
2345 if (nr_fair_skipped) {
2346 zonelist_rescan = true;
2347 reset_alloc_batches(ac->preferred_zone);
2349 if (nr_online_nodes > 1)
2350 zonelist_rescan = true;
2353 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2354 /* Disable zlc cache for second zonelist scan */
2356 zonelist_rescan = true;
2359 if (zonelist_rescan)
2366 * Large machines with many possible nodes should not always dump per-node
2367 * meminfo in irq context.
2369 static inline bool should_suppress_show_mem(void)
2374 ret = in_interrupt();
2379 static DEFINE_RATELIMIT_STATE(nopage_rs,
2380 DEFAULT_RATELIMIT_INTERVAL,
2381 DEFAULT_RATELIMIT_BURST);
2383 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2385 unsigned int filter = SHOW_MEM_FILTER_NODES;
2387 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2388 debug_guardpage_minorder() > 0)
2392 * This documents exceptions given to allocations in certain
2393 * contexts that are allowed to allocate outside current's set
2396 if (!(gfp_mask & __GFP_NOMEMALLOC))
2397 if (test_thread_flag(TIF_MEMDIE) ||
2398 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2399 filter &= ~SHOW_MEM_FILTER_NODES;
2400 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2401 filter &= ~SHOW_MEM_FILTER_NODES;
2404 struct va_format vaf;
2407 va_start(args, fmt);
2412 pr_warn("%pV", &vaf);
2417 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2418 current->comm, order, gfp_mask);
2421 if (!should_suppress_show_mem())
2425 static inline struct page *
2426 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2427 const struct alloc_context *ac, unsigned long *did_some_progress)
2431 *did_some_progress = 0;
2434 * Acquire the oom lock. If that fails, somebody else is
2435 * making progress for us.
2437 if (!mutex_trylock(&oom_lock)) {
2438 *did_some_progress = 1;
2439 schedule_timeout_uninterruptible(1);
2444 * Go through the zonelist yet one more time, keep very high watermark
2445 * here, this is only to catch a parallel oom killing, we must fail if
2446 * we're still under heavy pressure.
2448 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2449 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2453 if (!(gfp_mask & __GFP_NOFAIL)) {
2454 /* Coredumps can quickly deplete all memory reserves */
2455 if (current->flags & PF_DUMPCORE)
2457 /* The OOM killer will not help higher order allocs */
2458 if (order > PAGE_ALLOC_COSTLY_ORDER)
2460 /* The OOM killer does not needlessly kill tasks for lowmem */
2461 if (ac->high_zoneidx < ZONE_NORMAL)
2463 /* The OOM killer does not compensate for IO-less reclaim */
2464 if (!(gfp_mask & __GFP_FS)) {
2466 * XXX: Page reclaim didn't yield anything,
2467 * and the OOM killer can't be invoked, but
2468 * keep looping as per tradition.
2470 *did_some_progress = 1;
2473 if (pm_suspended_storage())
2475 /* The OOM killer may not free memory on a specific node */
2476 if (gfp_mask & __GFP_THISNODE)
2479 /* Exhausted what can be done so it's blamo time */
2480 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2481 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2482 *did_some_progress = 1;
2484 mutex_unlock(&oom_lock);
2488 #ifdef CONFIG_COMPACTION
2489 /* Try memory compaction for high-order allocations before reclaim */
2490 static struct page *
2491 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2492 int alloc_flags, const struct alloc_context *ac,
2493 enum migrate_mode mode, int *contended_compaction,
2494 bool *deferred_compaction)
2496 unsigned long compact_result;
2502 current->flags |= PF_MEMALLOC;
2503 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2504 mode, contended_compaction);
2505 current->flags &= ~PF_MEMALLOC;
2507 switch (compact_result) {
2508 case COMPACT_DEFERRED:
2509 *deferred_compaction = true;
2511 case COMPACT_SKIPPED:
2518 * At least in one zone compaction wasn't deferred or skipped, so let's
2519 * count a compaction stall
2521 count_vm_event(COMPACTSTALL);
2523 page = get_page_from_freelist(gfp_mask, order,
2524 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2527 struct zone *zone = page_zone(page);
2529 zone->compact_blockskip_flush = false;
2530 compaction_defer_reset(zone, order, true);
2531 count_vm_event(COMPACTSUCCESS);
2536 * It's bad if compaction run occurs and fails. The most likely reason
2537 * is that pages exist, but not enough to satisfy watermarks.
2539 count_vm_event(COMPACTFAIL);
2546 static inline struct page *
2547 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2548 int alloc_flags, const struct alloc_context *ac,
2549 enum migrate_mode mode, int *contended_compaction,
2550 bool *deferred_compaction)
2554 #endif /* CONFIG_COMPACTION */
2556 /* Perform direct synchronous page reclaim */
2558 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2559 const struct alloc_context *ac)
2561 struct reclaim_state reclaim_state;
2566 /* We now go into synchronous reclaim */
2567 cpuset_memory_pressure_bump();
2568 current->flags |= PF_MEMALLOC;
2569 lockdep_set_current_reclaim_state(gfp_mask);
2570 reclaim_state.reclaimed_slab = 0;
2571 current->reclaim_state = &reclaim_state;
2573 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2576 current->reclaim_state = NULL;
2577 lockdep_clear_current_reclaim_state();
2578 current->flags &= ~PF_MEMALLOC;
2585 /* The really slow allocator path where we enter direct reclaim */
2586 static inline struct page *
2587 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2588 int alloc_flags, const struct alloc_context *ac,
2589 unsigned long *did_some_progress)
2591 struct page *page = NULL;
2592 bool drained = false;
2594 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2595 if (unlikely(!(*did_some_progress)))
2598 /* After successful reclaim, reconsider all zones for allocation */
2599 if (IS_ENABLED(CONFIG_NUMA))
2600 zlc_clear_zones_full(ac->zonelist);
2603 page = get_page_from_freelist(gfp_mask, order,
2604 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2607 * If an allocation failed after direct reclaim, it could be because
2608 * pages are pinned on the per-cpu lists. Drain them and try again
2610 if (!page && !drained) {
2611 drain_all_pages(NULL);
2620 * This is called in the allocator slow-path if the allocation request is of
2621 * sufficient urgency to ignore watermarks and take other desperate measures
2623 static inline struct page *
2624 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2625 const struct alloc_context *ac)
2630 page = get_page_from_freelist(gfp_mask, order,
2631 ALLOC_NO_WATERMARKS, ac);
2633 if (!page && gfp_mask & __GFP_NOFAIL)
2634 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2636 } while (!page && (gfp_mask & __GFP_NOFAIL));
2641 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2646 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2647 ac->high_zoneidx, ac->nodemask)
2648 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2652 gfp_to_alloc_flags(gfp_t gfp_mask)
2654 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2655 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2657 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2658 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2661 * The caller may dip into page reserves a bit more if the caller
2662 * cannot run direct reclaim, or if the caller has realtime scheduling
2663 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2664 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2666 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2670 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2671 * if it can't schedule.
2673 if (!(gfp_mask & __GFP_NOMEMALLOC))
2674 alloc_flags |= ALLOC_HARDER;
2676 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2677 * comment for __cpuset_node_allowed().
2679 alloc_flags &= ~ALLOC_CPUSET;
2680 } else if (unlikely(rt_task(current)) && !in_interrupt())
2681 alloc_flags |= ALLOC_HARDER;
2683 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2684 if (gfp_mask & __GFP_MEMALLOC)
2685 alloc_flags |= ALLOC_NO_WATERMARKS;
2686 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2687 alloc_flags |= ALLOC_NO_WATERMARKS;
2688 else if (!in_interrupt() &&
2689 ((current->flags & PF_MEMALLOC) ||
2690 unlikely(test_thread_flag(TIF_MEMDIE))))
2691 alloc_flags |= ALLOC_NO_WATERMARKS;
2694 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2695 alloc_flags |= ALLOC_CMA;
2700 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2702 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2705 static inline struct page *
2706 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2707 struct alloc_context *ac)
2709 const gfp_t wait = gfp_mask & __GFP_WAIT;
2710 struct page *page = NULL;
2712 unsigned long pages_reclaimed = 0;
2713 unsigned long did_some_progress;
2714 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2715 bool deferred_compaction = false;
2716 int contended_compaction = COMPACT_CONTENDED_NONE;
2719 * In the slowpath, we sanity check order to avoid ever trying to
2720 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2721 * be using allocators in order of preference for an area that is
2724 if (order >= MAX_ORDER) {
2725 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2730 * If this allocation cannot block and it is for a specific node, then
2731 * fail early. There's no need to wakeup kswapd or retry for a
2732 * speculative node-specific allocation.
2734 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2738 if (!(gfp_mask & __GFP_NO_KSWAPD))
2739 wake_all_kswapds(order, ac);
2742 * OK, we're below the kswapd watermark and have kicked background
2743 * reclaim. Now things get more complex, so set up alloc_flags according
2744 * to how we want to proceed.
2746 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2749 * Find the true preferred zone if the allocation is unconstrained by
2752 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2753 struct zoneref *preferred_zoneref;
2754 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2755 ac->high_zoneidx, NULL, &ac->preferred_zone);
2756 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2759 /* This is the last chance, in general, before the goto nopage. */
2760 page = get_page_from_freelist(gfp_mask, order,
2761 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2765 /* Allocate without watermarks if the context allows */
2766 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2768 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2769 * the allocation is high priority and these type of
2770 * allocations are system rather than user orientated
2772 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2774 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2781 /* Atomic allocations - we can't balance anything */
2784 * All existing users of the deprecated __GFP_NOFAIL are
2785 * blockable, so warn of any new users that actually allow this
2786 * type of allocation to fail.
2788 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2792 /* Avoid recursion of direct reclaim */
2793 if (current->flags & PF_MEMALLOC)
2796 /* Avoid allocations with no watermarks from looping endlessly */
2797 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2801 * Try direct compaction. The first pass is asynchronous. Subsequent
2802 * attempts after direct reclaim are synchronous
2804 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2806 &contended_compaction,
2807 &deferred_compaction);
2811 /* Checks for THP-specific high-order allocations */
2812 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2814 * If compaction is deferred for high-order allocations, it is
2815 * because sync compaction recently failed. If this is the case
2816 * and the caller requested a THP allocation, we do not want
2817 * to heavily disrupt the system, so we fail the allocation
2818 * instead of entering direct reclaim.
2820 if (deferred_compaction)
2824 * In all zones where compaction was attempted (and not
2825 * deferred or skipped), lock contention has been detected.
2826 * For THP allocation we do not want to disrupt the others
2827 * so we fallback to base pages instead.
2829 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2833 * If compaction was aborted due to need_resched(), we do not
2834 * want to further increase allocation latency, unless it is
2835 * khugepaged trying to collapse.
2837 if (contended_compaction == COMPACT_CONTENDED_SCHED
2838 && !(current->flags & PF_KTHREAD))
2843 * It can become very expensive to allocate transparent hugepages at
2844 * fault, so use asynchronous memory compaction for THP unless it is
2845 * khugepaged trying to collapse.
2847 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2848 (current->flags & PF_KTHREAD))
2849 migration_mode = MIGRATE_SYNC_LIGHT;
2851 /* Try direct reclaim and then allocating */
2852 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2853 &did_some_progress);
2857 /* Do not loop if specifically requested */
2858 if (gfp_mask & __GFP_NORETRY)
2861 /* Keep reclaiming pages as long as there is reasonable progress */
2862 pages_reclaimed += did_some_progress;
2863 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
2864 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
2865 /* Wait for some write requests to complete then retry */
2866 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2870 /* Reclaim has failed us, start killing things */
2871 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
2875 /* Retry as long as the OOM killer is making progress */
2876 if (did_some_progress)
2881 * High-order allocations do not necessarily loop after
2882 * direct reclaim and reclaim/compaction depends on compaction
2883 * being called after reclaim so call directly if necessary
2885 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
2887 &contended_compaction,
2888 &deferred_compaction);
2892 warn_alloc_failed(gfp_mask, order, NULL);
2898 * This is the 'heart' of the zoned buddy allocator.
2901 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2902 struct zonelist *zonelist, nodemask_t *nodemask)
2904 struct zoneref *preferred_zoneref;
2905 struct page *page = NULL;
2906 unsigned int cpuset_mems_cookie;
2907 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2908 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2909 struct alloc_context ac = {
2910 .high_zoneidx = gfp_zone(gfp_mask),
2911 .nodemask = nodemask,
2912 .migratetype = gfpflags_to_migratetype(gfp_mask),
2915 gfp_mask &= gfp_allowed_mask;
2917 lockdep_trace_alloc(gfp_mask);
2919 might_sleep_if(gfp_mask & __GFP_WAIT);
2921 if (should_fail_alloc_page(gfp_mask, order))
2925 * Check the zones suitable for the gfp_mask contain at least one
2926 * valid zone. It's possible to have an empty zonelist as a result
2927 * of __GFP_THISNODE and a memoryless node
2929 if (unlikely(!zonelist->_zonerefs->zone))
2932 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2933 alloc_flags |= ALLOC_CMA;
2936 cpuset_mems_cookie = read_mems_allowed_begin();
2938 /* We set it here, as __alloc_pages_slowpath might have changed it */
2939 ac.zonelist = zonelist;
2940 /* The preferred zone is used for statistics later */
2941 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2942 ac.nodemask ? : &cpuset_current_mems_allowed,
2943 &ac.preferred_zone);
2944 if (!ac.preferred_zone)
2946 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2948 /* First allocation attempt */
2949 alloc_mask = gfp_mask|__GFP_HARDWALL;
2950 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2951 if (unlikely(!page)) {
2953 * Runtime PM, block IO and its error handling path
2954 * can deadlock because I/O on the device might not
2957 alloc_mask = memalloc_noio_flags(gfp_mask);
2959 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2962 if (kmemcheck_enabled && page)
2963 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2965 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2969 * When updating a task's mems_allowed, it is possible to race with
2970 * parallel threads in such a way that an allocation can fail while
2971 * the mask is being updated. If a page allocation is about to fail,
2972 * check if the cpuset changed during allocation and if so, retry.
2974 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2979 EXPORT_SYMBOL(__alloc_pages_nodemask);
2982 * Common helper functions.
2984 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2989 * __get_free_pages() returns a 32-bit address, which cannot represent
2992 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2994 page = alloc_pages(gfp_mask, order);
2997 return (unsigned long) page_address(page);
2999 EXPORT_SYMBOL(__get_free_pages);
3001 unsigned long get_zeroed_page(gfp_t gfp_mask)
3003 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3005 EXPORT_SYMBOL(get_zeroed_page);
3007 void __free_pages(struct page *page, unsigned int order)
3009 if (put_page_testzero(page)) {
3011 free_hot_cold_page(page, false);
3013 __free_pages_ok(page, order);
3017 EXPORT_SYMBOL(__free_pages);
3019 void free_pages(unsigned long addr, unsigned int order)
3022 VM_BUG_ON(!virt_addr_valid((void *)addr));
3023 __free_pages(virt_to_page((void *)addr), order);
3027 EXPORT_SYMBOL(free_pages);
3031 * An arbitrary-length arbitrary-offset area of memory which resides
3032 * within a 0 or higher order page. Multiple fragments within that page
3033 * are individually refcounted, in the page's reference counter.
3035 * The page_frag functions below provide a simple allocation framework for
3036 * page fragments. This is used by the network stack and network device
3037 * drivers to provide a backing region of memory for use as either an
3038 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3040 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3043 struct page *page = NULL;
3044 gfp_t gfp = gfp_mask;
3046 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3047 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3049 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3050 PAGE_FRAG_CACHE_MAX_ORDER);
3051 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3053 if (unlikely(!page))
3054 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3056 nc->va = page ? page_address(page) : NULL;
3061 void *__alloc_page_frag(struct page_frag_cache *nc,
3062 unsigned int fragsz, gfp_t gfp_mask)
3064 unsigned int size = PAGE_SIZE;
3068 if (unlikely(!nc->va)) {
3070 page = __page_frag_refill(nc, gfp_mask);
3074 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3075 /* if size can vary use size else just use PAGE_SIZE */
3078 /* Even if we own the page, we do not use atomic_set().
3079 * This would break get_page_unless_zero() users.
3081 atomic_add(size - 1, &page->_count);
3083 /* reset page count bias and offset to start of new frag */
3084 nc->pfmemalloc = page->pfmemalloc;
3085 nc->pagecnt_bias = size;
3089 offset = nc->offset - fragsz;
3090 if (unlikely(offset < 0)) {
3091 page = virt_to_page(nc->va);
3093 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3096 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3097 /* if size can vary use size else just use PAGE_SIZE */
3100 /* OK, page count is 0, we can safely set it */
3101 atomic_set(&page->_count, size);
3103 /* reset page count bias and offset to start of new frag */
3104 nc->pagecnt_bias = size;
3105 offset = size - fragsz;
3109 nc->offset = offset;
3111 return nc->va + offset;
3113 EXPORT_SYMBOL(__alloc_page_frag);
3116 * Frees a page fragment allocated out of either a compound or order 0 page.
3118 void __free_page_frag(void *addr)
3120 struct page *page = virt_to_head_page(addr);
3122 if (unlikely(put_page_testzero(page)))
3123 __free_pages_ok(page, compound_order(page));
3125 EXPORT_SYMBOL(__free_page_frag);
3128 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3129 * of the current memory cgroup.
3131 * It should be used when the caller would like to use kmalloc, but since the
3132 * allocation is large, it has to fall back to the page allocator.
3134 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3137 struct mem_cgroup *memcg = NULL;
3139 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3141 page = alloc_pages(gfp_mask, order);
3142 memcg_kmem_commit_charge(page, memcg, order);
3146 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3149 struct mem_cgroup *memcg = NULL;
3151 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3153 page = alloc_pages_node(nid, gfp_mask, order);
3154 memcg_kmem_commit_charge(page, memcg, order);
3159 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3162 void __free_kmem_pages(struct page *page, unsigned int order)
3164 memcg_kmem_uncharge_pages(page, order);
3165 __free_pages(page, order);
3168 void free_kmem_pages(unsigned long addr, unsigned int order)
3171 VM_BUG_ON(!virt_addr_valid((void *)addr));
3172 __free_kmem_pages(virt_to_page((void *)addr), order);
3176 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3179 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3180 unsigned long used = addr + PAGE_ALIGN(size);
3182 split_page(virt_to_page((void *)addr), order);
3183 while (used < alloc_end) {
3188 return (void *)addr;
3192 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3193 * @size: the number of bytes to allocate
3194 * @gfp_mask: GFP flags for the allocation
3196 * This function is similar to alloc_pages(), except that it allocates the
3197 * minimum number of pages to satisfy the request. alloc_pages() can only
3198 * allocate memory in power-of-two pages.
3200 * This function is also limited by MAX_ORDER.
3202 * Memory allocated by this function must be released by free_pages_exact().
3204 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3206 unsigned int order = get_order(size);
3209 addr = __get_free_pages(gfp_mask, order);
3210 return make_alloc_exact(addr, order, size);
3212 EXPORT_SYMBOL(alloc_pages_exact);
3215 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3217 * @nid: the preferred node ID where memory should be allocated
3218 * @size: the number of bytes to allocate
3219 * @gfp_mask: GFP flags for the allocation
3221 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3223 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3226 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3228 unsigned order = get_order(size);
3229 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3232 return make_alloc_exact((unsigned long)page_address(p), order, size);
3236 * free_pages_exact - release memory allocated via alloc_pages_exact()
3237 * @virt: the value returned by alloc_pages_exact.
3238 * @size: size of allocation, same value as passed to alloc_pages_exact().
3240 * Release the memory allocated by a previous call to alloc_pages_exact.
3242 void free_pages_exact(void *virt, size_t size)
3244 unsigned long addr = (unsigned long)virt;
3245 unsigned long end = addr + PAGE_ALIGN(size);
3247 while (addr < end) {
3252 EXPORT_SYMBOL(free_pages_exact);
3255 * nr_free_zone_pages - count number of pages beyond high watermark
3256 * @offset: The zone index of the highest zone
3258 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3259 * high watermark within all zones at or below a given zone index. For each
3260 * zone, the number of pages is calculated as:
3261 * managed_pages - high_pages
3263 static unsigned long nr_free_zone_pages(int offset)
3268 /* Just pick one node, since fallback list is circular */
3269 unsigned long sum = 0;
3271 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3273 for_each_zone_zonelist(zone, z, zonelist, offset) {
3274 unsigned long size = zone->managed_pages;
3275 unsigned long high = high_wmark_pages(zone);
3284 * nr_free_buffer_pages - count number of pages beyond high watermark
3286 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3287 * watermark within ZONE_DMA and ZONE_NORMAL.
3289 unsigned long nr_free_buffer_pages(void)
3291 return nr_free_zone_pages(gfp_zone(GFP_USER));
3293 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3296 * nr_free_pagecache_pages - count number of pages beyond high watermark
3298 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3299 * high watermark within all zones.
3301 unsigned long nr_free_pagecache_pages(void)
3303 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3306 static inline void show_node(struct zone *zone)
3308 if (IS_ENABLED(CONFIG_NUMA))
3309 printk("Node %d ", zone_to_nid(zone));
3312 void si_meminfo(struct sysinfo *val)
3314 val->totalram = totalram_pages;
3315 val->sharedram = global_page_state(NR_SHMEM);
3316 val->freeram = global_page_state(NR_FREE_PAGES);
3317 val->bufferram = nr_blockdev_pages();
3318 val->totalhigh = totalhigh_pages;
3319 val->freehigh = nr_free_highpages();
3320 val->mem_unit = PAGE_SIZE;
3323 EXPORT_SYMBOL(si_meminfo);
3326 void si_meminfo_node(struct sysinfo *val, int nid)
3328 int zone_type; /* needs to be signed */
3329 unsigned long managed_pages = 0;
3330 pg_data_t *pgdat = NODE_DATA(nid);
3332 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3333 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3334 val->totalram = managed_pages;
3335 val->sharedram = node_page_state(nid, NR_SHMEM);
3336 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3337 #ifdef CONFIG_HIGHMEM
3338 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3339 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3345 val->mem_unit = PAGE_SIZE;
3350 * Determine whether the node should be displayed or not, depending on whether
3351 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3353 bool skip_free_areas_node(unsigned int flags, int nid)
3356 unsigned int cpuset_mems_cookie;
3358 if (!(flags & SHOW_MEM_FILTER_NODES))
3362 cpuset_mems_cookie = read_mems_allowed_begin();
3363 ret = !node_isset(nid, cpuset_current_mems_allowed);
3364 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3369 #define K(x) ((x) << (PAGE_SHIFT-10))
3371 static void show_migration_types(unsigned char type)
3373 static const char types[MIGRATE_TYPES] = {
3374 [MIGRATE_UNMOVABLE] = 'U',
3375 [MIGRATE_RECLAIMABLE] = 'E',
3376 [MIGRATE_MOVABLE] = 'M',
3377 [MIGRATE_RESERVE] = 'R',
3379 [MIGRATE_CMA] = 'C',
3381 #ifdef CONFIG_MEMORY_ISOLATION
3382 [MIGRATE_ISOLATE] = 'I',
3385 char tmp[MIGRATE_TYPES + 1];
3389 for (i = 0; i < MIGRATE_TYPES; i++) {
3390 if (type & (1 << i))
3395 printk("(%s) ", tmp);
3399 * Show free area list (used inside shift_scroll-lock stuff)
3400 * We also calculate the percentage fragmentation. We do this by counting the
3401 * memory on each free list with the exception of the first item on the list.
3404 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3407 void show_free_areas(unsigned int filter)
3409 unsigned long free_pcp = 0;
3413 for_each_populated_zone(zone) {
3414 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3417 for_each_online_cpu(cpu)
3418 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3421 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3422 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3423 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3424 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3425 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3426 " free:%lu free_pcp:%lu free_cma:%lu\n",
3427 global_page_state(NR_ACTIVE_ANON),
3428 global_page_state(NR_INACTIVE_ANON),
3429 global_page_state(NR_ISOLATED_ANON),
3430 global_page_state(NR_ACTIVE_FILE),
3431 global_page_state(NR_INACTIVE_FILE),
3432 global_page_state(NR_ISOLATED_FILE),
3433 global_page_state(NR_UNEVICTABLE),
3434 global_page_state(NR_FILE_DIRTY),
3435 global_page_state(NR_WRITEBACK),
3436 global_page_state(NR_UNSTABLE_NFS),
3437 global_page_state(NR_SLAB_RECLAIMABLE),
3438 global_page_state(NR_SLAB_UNRECLAIMABLE),
3439 global_page_state(NR_FILE_MAPPED),
3440 global_page_state(NR_SHMEM),
3441 global_page_state(NR_PAGETABLE),
3442 global_page_state(NR_BOUNCE),
3443 global_page_state(NR_FREE_PAGES),
3445 global_page_state(NR_FREE_CMA_PAGES));
3447 for_each_populated_zone(zone) {
3450 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3454 for_each_online_cpu(cpu)
3455 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3463 " active_anon:%lukB"
3464 " inactive_anon:%lukB"
3465 " active_file:%lukB"
3466 " inactive_file:%lukB"
3467 " unevictable:%lukB"
3468 " isolated(anon):%lukB"
3469 " isolated(file):%lukB"
3477 " slab_reclaimable:%lukB"
3478 " slab_unreclaimable:%lukB"
3479 " kernel_stack:%lukB"
3486 " writeback_tmp:%lukB"
3487 " pages_scanned:%lu"
3488 " all_unreclaimable? %s"
3491 K(zone_page_state(zone, NR_FREE_PAGES)),
3492 K(min_wmark_pages(zone)),
3493 K(low_wmark_pages(zone)),
3494 K(high_wmark_pages(zone)),
3495 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3496 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3497 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3498 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3499 K(zone_page_state(zone, NR_UNEVICTABLE)),
3500 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3501 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3502 K(zone->present_pages),
3503 K(zone->managed_pages),
3504 K(zone_page_state(zone, NR_MLOCK)),
3505 K(zone_page_state(zone, NR_FILE_DIRTY)),
3506 K(zone_page_state(zone, NR_WRITEBACK)),
3507 K(zone_page_state(zone, NR_FILE_MAPPED)),
3508 K(zone_page_state(zone, NR_SHMEM)),
3509 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3510 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3511 zone_page_state(zone, NR_KERNEL_STACK) *
3513 K(zone_page_state(zone, NR_PAGETABLE)),
3514 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3515 K(zone_page_state(zone, NR_BOUNCE)),
3517 K(this_cpu_read(zone->pageset->pcp.count)),
3518 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3519 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3520 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3521 (!zone_reclaimable(zone) ? "yes" : "no")
3523 printk("lowmem_reserve[]:");
3524 for (i = 0; i < MAX_NR_ZONES; i++)
3525 printk(" %ld", zone->lowmem_reserve[i]);
3529 for_each_populated_zone(zone) {
3530 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3531 unsigned char types[MAX_ORDER];
3533 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3536 printk("%s: ", zone->name);
3538 spin_lock_irqsave(&zone->lock, flags);
3539 for (order = 0; order < MAX_ORDER; order++) {
3540 struct free_area *area = &zone->free_area[order];
3543 nr[order] = area->nr_free;
3544 total += nr[order] << order;
3547 for (type = 0; type < MIGRATE_TYPES; type++) {
3548 if (!list_empty(&area->free_list[type]))
3549 types[order] |= 1 << type;
3552 spin_unlock_irqrestore(&zone->lock, flags);
3553 for (order = 0; order < MAX_ORDER; order++) {
3554 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3556 show_migration_types(types[order]);
3558 printk("= %lukB\n", K(total));
3561 hugetlb_show_meminfo();
3563 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3565 show_swap_cache_info();
3568 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3570 zoneref->zone = zone;
3571 zoneref->zone_idx = zone_idx(zone);
3575 * Builds allocation fallback zone lists.
3577 * Add all populated zones of a node to the zonelist.
3579 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3583 enum zone_type zone_type = MAX_NR_ZONES;
3587 zone = pgdat->node_zones + zone_type;
3588 if (populated_zone(zone)) {
3589 zoneref_set_zone(zone,
3590 &zonelist->_zonerefs[nr_zones++]);
3591 check_highest_zone(zone_type);
3593 } while (zone_type);
3601 * 0 = automatic detection of better ordering.
3602 * 1 = order by ([node] distance, -zonetype)
3603 * 2 = order by (-zonetype, [node] distance)
3605 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3606 * the same zonelist. So only NUMA can configure this param.
3608 #define ZONELIST_ORDER_DEFAULT 0
3609 #define ZONELIST_ORDER_NODE 1
3610 #define ZONELIST_ORDER_ZONE 2
3612 /* zonelist order in the kernel.
3613 * set_zonelist_order() will set this to NODE or ZONE.
3615 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3616 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3620 /* The value user specified ....changed by config */
3621 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3622 /* string for sysctl */
3623 #define NUMA_ZONELIST_ORDER_LEN 16
3624 char numa_zonelist_order[16] = "default";
3627 * interface for configure zonelist ordering.
3628 * command line option "numa_zonelist_order"
3629 * = "[dD]efault - default, automatic configuration.
3630 * = "[nN]ode - order by node locality, then by zone within node
3631 * = "[zZ]one - order by zone, then by locality within zone
3634 static int __parse_numa_zonelist_order(char *s)
3636 if (*s == 'd' || *s == 'D') {
3637 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3638 } else if (*s == 'n' || *s == 'N') {
3639 user_zonelist_order = ZONELIST_ORDER_NODE;
3640 } else if (*s == 'z' || *s == 'Z') {
3641 user_zonelist_order = ZONELIST_ORDER_ZONE;
3644 "Ignoring invalid numa_zonelist_order value: "
3651 static __init int setup_numa_zonelist_order(char *s)
3658 ret = __parse_numa_zonelist_order(s);
3660 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3664 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3667 * sysctl handler for numa_zonelist_order
3669 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3670 void __user *buffer, size_t *length,
3673 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3675 static DEFINE_MUTEX(zl_order_mutex);
3677 mutex_lock(&zl_order_mutex);
3679 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3683 strcpy(saved_string, (char *)table->data);
3685 ret = proc_dostring(table, write, buffer, length, ppos);
3689 int oldval = user_zonelist_order;
3691 ret = __parse_numa_zonelist_order((char *)table->data);
3694 * bogus value. restore saved string
3696 strncpy((char *)table->data, saved_string,
3697 NUMA_ZONELIST_ORDER_LEN);
3698 user_zonelist_order = oldval;
3699 } else if (oldval != user_zonelist_order) {
3700 mutex_lock(&zonelists_mutex);
3701 build_all_zonelists(NULL, NULL);
3702 mutex_unlock(&zonelists_mutex);
3706 mutex_unlock(&zl_order_mutex);
3711 #define MAX_NODE_LOAD (nr_online_nodes)
3712 static int node_load[MAX_NUMNODES];
3715 * find_next_best_node - find the next node that should appear in a given node's fallback list
3716 * @node: node whose fallback list we're appending
3717 * @used_node_mask: nodemask_t of already used nodes
3719 * We use a number of factors to determine which is the next node that should
3720 * appear on a given node's fallback list. The node should not have appeared
3721 * already in @node's fallback list, and it should be the next closest node
3722 * according to the distance array (which contains arbitrary distance values
3723 * from each node to each node in the system), and should also prefer nodes
3724 * with no CPUs, since presumably they'll have very little allocation pressure
3725 * on them otherwise.
3726 * It returns -1 if no node is found.
3728 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3731 int min_val = INT_MAX;
3732 int best_node = NUMA_NO_NODE;
3733 const struct cpumask *tmp = cpumask_of_node(0);
3735 /* Use the local node if we haven't already */
3736 if (!node_isset(node, *used_node_mask)) {
3737 node_set(node, *used_node_mask);
3741 for_each_node_state(n, N_MEMORY) {
3743 /* Don't want a node to appear more than once */
3744 if (node_isset(n, *used_node_mask))
3747 /* Use the distance array to find the distance */
3748 val = node_distance(node, n);
3750 /* Penalize nodes under us ("prefer the next node") */
3753 /* Give preference to headless and unused nodes */
3754 tmp = cpumask_of_node(n);
3755 if (!cpumask_empty(tmp))
3756 val += PENALTY_FOR_NODE_WITH_CPUS;
3758 /* Slight preference for less loaded node */
3759 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3760 val += node_load[n];
3762 if (val < min_val) {
3769 node_set(best_node, *used_node_mask);
3776 * Build zonelists ordered by node and zones within node.
3777 * This results in maximum locality--normal zone overflows into local
3778 * DMA zone, if any--but risks exhausting DMA zone.
3780 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3783 struct zonelist *zonelist;
3785 zonelist = &pgdat->node_zonelists[0];
3786 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3788 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3789 zonelist->_zonerefs[j].zone = NULL;
3790 zonelist->_zonerefs[j].zone_idx = 0;
3794 * Build gfp_thisnode zonelists
3796 static void build_thisnode_zonelists(pg_data_t *pgdat)
3799 struct zonelist *zonelist;
3801 zonelist = &pgdat->node_zonelists[1];
3802 j = build_zonelists_node(pgdat, zonelist, 0);
3803 zonelist->_zonerefs[j].zone = NULL;
3804 zonelist->_zonerefs[j].zone_idx = 0;
3808 * Build zonelists ordered by zone and nodes within zones.
3809 * This results in conserving DMA zone[s] until all Normal memory is
3810 * exhausted, but results in overflowing to remote node while memory
3811 * may still exist in local DMA zone.
3813 static int node_order[MAX_NUMNODES];
3815 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3818 int zone_type; /* needs to be signed */
3820 struct zonelist *zonelist;
3822 zonelist = &pgdat->node_zonelists[0];
3824 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3825 for (j = 0; j < nr_nodes; j++) {
3826 node = node_order[j];
3827 z = &NODE_DATA(node)->node_zones[zone_type];
3828 if (populated_zone(z)) {
3830 &zonelist->_zonerefs[pos++]);
3831 check_highest_zone(zone_type);
3835 zonelist->_zonerefs[pos].zone = NULL;
3836 zonelist->_zonerefs[pos].zone_idx = 0;
3839 #if defined(CONFIG_64BIT)
3841 * Devices that require DMA32/DMA are relatively rare and do not justify a
3842 * penalty to every machine in case the specialised case applies. Default
3843 * to Node-ordering on 64-bit NUMA machines
3845 static int default_zonelist_order(void)
3847 return ZONELIST_ORDER_NODE;
3851 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3852 * by the kernel. If processes running on node 0 deplete the low memory zone
3853 * then reclaim will occur more frequency increasing stalls and potentially
3854 * be easier to OOM if a large percentage of the zone is under writeback or
3855 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3856 * Hence, default to zone ordering on 32-bit.
3858 static int default_zonelist_order(void)
3860 return ZONELIST_ORDER_ZONE;
3862 #endif /* CONFIG_64BIT */
3864 static void set_zonelist_order(void)
3866 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3867 current_zonelist_order = default_zonelist_order();
3869 current_zonelist_order = user_zonelist_order;
3872 static void build_zonelists(pg_data_t *pgdat)
3876 nodemask_t used_mask;
3877 int local_node, prev_node;
3878 struct zonelist *zonelist;
3879 int order = current_zonelist_order;
3881 /* initialize zonelists */
3882 for (i = 0; i < MAX_ZONELISTS; i++) {
3883 zonelist = pgdat->node_zonelists + i;
3884 zonelist->_zonerefs[0].zone = NULL;
3885 zonelist->_zonerefs[0].zone_idx = 0;
3888 /* NUMA-aware ordering of nodes */
3889 local_node = pgdat->node_id;
3890 load = nr_online_nodes;
3891 prev_node = local_node;
3892 nodes_clear(used_mask);
3894 memset(node_order, 0, sizeof(node_order));
3897 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3899 * We don't want to pressure a particular node.
3900 * So adding penalty to the first node in same
3901 * distance group to make it round-robin.
3903 if (node_distance(local_node, node) !=
3904 node_distance(local_node, prev_node))
3905 node_load[node] = load;
3909 if (order == ZONELIST_ORDER_NODE)
3910 build_zonelists_in_node_order(pgdat, node);
3912 node_order[j++] = node; /* remember order */
3915 if (order == ZONELIST_ORDER_ZONE) {
3916 /* calculate node order -- i.e., DMA last! */
3917 build_zonelists_in_zone_order(pgdat, j);
3920 build_thisnode_zonelists(pgdat);
3923 /* Construct the zonelist performance cache - see further mmzone.h */
3924 static void build_zonelist_cache(pg_data_t *pgdat)
3926 struct zonelist *zonelist;
3927 struct zonelist_cache *zlc;
3930 zonelist = &pgdat->node_zonelists[0];
3931 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3932 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3933 for (z = zonelist->_zonerefs; z->zone; z++)
3934 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3937 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3939 * Return node id of node used for "local" allocations.
3940 * I.e., first node id of first zone in arg node's generic zonelist.
3941 * Used for initializing percpu 'numa_mem', which is used primarily
3942 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3944 int local_memory_node(int node)
3948 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3949 gfp_zone(GFP_KERNEL),
3956 #else /* CONFIG_NUMA */
3958 static void set_zonelist_order(void)
3960 current_zonelist_order = ZONELIST_ORDER_ZONE;
3963 static void build_zonelists(pg_data_t *pgdat)
3965 int node, local_node;
3967 struct zonelist *zonelist;
3969 local_node = pgdat->node_id;
3971 zonelist = &pgdat->node_zonelists[0];
3972 j = build_zonelists_node(pgdat, zonelist, 0);
3975 * Now we build the zonelist so that it contains the zones
3976 * of all the other nodes.
3977 * We don't want to pressure a particular node, so when
3978 * building the zones for node N, we make sure that the
3979 * zones coming right after the local ones are those from
3980 * node N+1 (modulo N)
3982 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3983 if (!node_online(node))
3985 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3987 for (node = 0; node < local_node; node++) {
3988 if (!node_online(node))
3990 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3993 zonelist->_zonerefs[j].zone = NULL;
3994 zonelist->_zonerefs[j].zone_idx = 0;
3997 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3998 static void build_zonelist_cache(pg_data_t *pgdat)
4000 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4003 #endif /* CONFIG_NUMA */
4006 * Boot pageset table. One per cpu which is going to be used for all
4007 * zones and all nodes. The parameters will be set in such a way
4008 * that an item put on a list will immediately be handed over to
4009 * the buddy list. This is safe since pageset manipulation is done
4010 * with interrupts disabled.
4012 * The boot_pagesets must be kept even after bootup is complete for
4013 * unused processors and/or zones. They do play a role for bootstrapping
4014 * hotplugged processors.
4016 * zoneinfo_show() and maybe other functions do
4017 * not check if the processor is online before following the pageset pointer.
4018 * Other parts of the kernel may not check if the zone is available.
4020 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4021 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4022 static void setup_zone_pageset(struct zone *zone);
4025 * Global mutex to protect against size modification of zonelists
4026 * as well as to serialize pageset setup for the new populated zone.
4028 DEFINE_MUTEX(zonelists_mutex);
4030 /* return values int ....just for stop_machine() */
4031 static int __build_all_zonelists(void *data)
4035 pg_data_t *self = data;
4038 memset(node_load, 0, sizeof(node_load));
4041 if (self && !node_online(self->node_id)) {
4042 build_zonelists(self);
4043 build_zonelist_cache(self);
4046 for_each_online_node(nid) {
4047 pg_data_t *pgdat = NODE_DATA(nid);
4049 build_zonelists(pgdat);
4050 build_zonelist_cache(pgdat);
4054 * Initialize the boot_pagesets that are going to be used
4055 * for bootstrapping processors. The real pagesets for
4056 * each zone will be allocated later when the per cpu
4057 * allocator is available.
4059 * boot_pagesets are used also for bootstrapping offline
4060 * cpus if the system is already booted because the pagesets
4061 * are needed to initialize allocators on a specific cpu too.
4062 * F.e. the percpu allocator needs the page allocator which
4063 * needs the percpu allocator in order to allocate its pagesets
4064 * (a chicken-egg dilemma).
4066 for_each_possible_cpu(cpu) {
4067 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4069 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4071 * We now know the "local memory node" for each node--
4072 * i.e., the node of the first zone in the generic zonelist.
4073 * Set up numa_mem percpu variable for on-line cpus. During
4074 * boot, only the boot cpu should be on-line; we'll init the
4075 * secondary cpus' numa_mem as they come on-line. During
4076 * node/memory hotplug, we'll fixup all on-line cpus.
4078 if (cpu_online(cpu))
4079 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4086 static noinline void __init
4087 build_all_zonelists_init(void)
4089 __build_all_zonelists(NULL);
4090 mminit_verify_zonelist();
4091 cpuset_init_current_mems_allowed();
4095 * Called with zonelists_mutex held always
4096 * unless system_state == SYSTEM_BOOTING.
4098 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4099 * [we're only called with non-NULL zone through __meminit paths] and
4100 * (2) call of __init annotated helper build_all_zonelists_init
4101 * [protected by SYSTEM_BOOTING].
4103 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4105 set_zonelist_order();
4107 if (system_state == SYSTEM_BOOTING) {
4108 build_all_zonelists_init();
4110 #ifdef CONFIG_MEMORY_HOTPLUG
4112 setup_zone_pageset(zone);
4114 /* we have to stop all cpus to guarantee there is no user
4116 stop_machine(__build_all_zonelists, pgdat, NULL);
4117 /* cpuset refresh routine should be here */
4119 vm_total_pages = nr_free_pagecache_pages();
4121 * Disable grouping by mobility if the number of pages in the
4122 * system is too low to allow the mechanism to work. It would be
4123 * more accurate, but expensive to check per-zone. This check is
4124 * made on memory-hotadd so a system can start with mobility
4125 * disabled and enable it later
4127 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4128 page_group_by_mobility_disabled = 1;
4130 page_group_by_mobility_disabled = 0;
4132 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4133 "Total pages: %ld\n",
4135 zonelist_order_name[current_zonelist_order],
4136 page_group_by_mobility_disabled ? "off" : "on",
4139 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4144 * Helper functions to size the waitqueue hash table.
4145 * Essentially these want to choose hash table sizes sufficiently
4146 * large so that collisions trying to wait on pages are rare.
4147 * But in fact, the number of active page waitqueues on typical
4148 * systems is ridiculously low, less than 200. So this is even
4149 * conservative, even though it seems large.
4151 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4152 * waitqueues, i.e. the size of the waitq table given the number of pages.
4154 #define PAGES_PER_WAITQUEUE 256
4156 #ifndef CONFIG_MEMORY_HOTPLUG
4157 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4159 unsigned long size = 1;
4161 pages /= PAGES_PER_WAITQUEUE;
4163 while (size < pages)
4167 * Once we have dozens or even hundreds of threads sleeping
4168 * on IO we've got bigger problems than wait queue collision.
4169 * Limit the size of the wait table to a reasonable size.
4171 size = min(size, 4096UL);
4173 return max(size, 4UL);
4177 * A zone's size might be changed by hot-add, so it is not possible to determine
4178 * a suitable size for its wait_table. So we use the maximum size now.
4180 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4182 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4183 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4184 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4186 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4187 * or more by the traditional way. (See above). It equals:
4189 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4190 * ia64(16K page size) : = ( 8G + 4M)byte.
4191 * powerpc (64K page size) : = (32G +16M)byte.
4193 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4200 * This is an integer logarithm so that shifts can be used later
4201 * to extract the more random high bits from the multiplicative
4202 * hash function before the remainder is taken.
4204 static inline unsigned long wait_table_bits(unsigned long size)
4210 * Check if a pageblock contains reserved pages
4212 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4216 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4217 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4224 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4225 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4226 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4227 * higher will lead to a bigger reserve which will get freed as contiguous
4228 * blocks as reclaim kicks in
4230 static void setup_zone_migrate_reserve(struct zone *zone)
4232 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4234 unsigned long block_migratetype;
4239 * Get the start pfn, end pfn and the number of blocks to reserve
4240 * We have to be careful to be aligned to pageblock_nr_pages to
4241 * make sure that we always check pfn_valid for the first page in
4244 start_pfn = zone->zone_start_pfn;
4245 end_pfn = zone_end_pfn(zone);
4246 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4247 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4251 * Reserve blocks are generally in place to help high-order atomic
4252 * allocations that are short-lived. A min_free_kbytes value that
4253 * would result in more than 2 reserve blocks for atomic allocations
4254 * is assumed to be in place to help anti-fragmentation for the
4255 * future allocation of hugepages at runtime.
4257 reserve = min(2, reserve);
4258 old_reserve = zone->nr_migrate_reserve_block;
4260 /* When memory hot-add, we almost always need to do nothing */
4261 if (reserve == old_reserve)
4263 zone->nr_migrate_reserve_block = reserve;
4265 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4266 if (!pfn_valid(pfn))
4268 page = pfn_to_page(pfn);
4270 /* Watch out for overlapping nodes */
4271 if (page_to_nid(page) != zone_to_nid(zone))
4274 block_migratetype = get_pageblock_migratetype(page);
4276 /* Only test what is necessary when the reserves are not met */
4279 * Blocks with reserved pages will never free, skip
4282 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4283 if (pageblock_is_reserved(pfn, block_end_pfn))
4286 /* If this block is reserved, account for it */
4287 if (block_migratetype == MIGRATE_RESERVE) {
4292 /* Suitable for reserving if this block is movable */
4293 if (block_migratetype == MIGRATE_MOVABLE) {
4294 set_pageblock_migratetype(page,
4296 move_freepages_block(zone, page,
4301 } else if (!old_reserve) {
4303 * At boot time we don't need to scan the whole zone
4304 * for turning off MIGRATE_RESERVE.
4310 * If the reserve is met and this is a previous reserved block,
4313 if (block_migratetype == MIGRATE_RESERVE) {
4314 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4315 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4321 * Initially all pages are reserved - free ones are freed
4322 * up by free_all_bootmem() once the early boot process is
4323 * done. Non-atomic initialization, single-pass.
4325 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4326 unsigned long start_pfn, enum memmap_context context)
4328 unsigned long end_pfn = start_pfn + size;
4332 if (highest_memmap_pfn < end_pfn - 1)
4333 highest_memmap_pfn = end_pfn - 1;
4335 z = &NODE_DATA(nid)->node_zones[zone];
4336 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4338 * There can be holes in boot-time mem_map[]s
4339 * handed to this function. They do not
4340 * exist on hotplugged memory.
4342 if (context == MEMMAP_EARLY) {
4343 if (!early_pfn_valid(pfn))
4345 if (!early_pfn_in_nid(pfn, nid))
4348 __init_single_pfn(pfn, zone, nid);
4352 static void __meminit zone_init_free_lists(struct zone *zone)
4354 unsigned int order, t;
4355 for_each_migratetype_order(order, t) {
4356 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4357 zone->free_area[order].nr_free = 0;
4361 #ifndef __HAVE_ARCH_MEMMAP_INIT
4362 #define memmap_init(size, nid, zone, start_pfn) \
4363 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4366 static int zone_batchsize(struct zone *zone)
4372 * The per-cpu-pages pools are set to around 1000th of the
4373 * size of the zone. But no more than 1/2 of a meg.
4375 * OK, so we don't know how big the cache is. So guess.
4377 batch = zone->managed_pages / 1024;
4378 if (batch * PAGE_SIZE > 512 * 1024)
4379 batch = (512 * 1024) / PAGE_SIZE;
4380 batch /= 4; /* We effectively *= 4 below */
4385 * Clamp the batch to a 2^n - 1 value. Having a power
4386 * of 2 value was found to be more likely to have
4387 * suboptimal cache aliasing properties in some cases.
4389 * For example if 2 tasks are alternately allocating
4390 * batches of pages, one task can end up with a lot
4391 * of pages of one half of the possible page colors
4392 * and the other with pages of the other colors.
4394 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4399 /* The deferral and batching of frees should be suppressed under NOMMU
4402 * The problem is that NOMMU needs to be able to allocate large chunks
4403 * of contiguous memory as there's no hardware page translation to
4404 * assemble apparent contiguous memory from discontiguous pages.
4406 * Queueing large contiguous runs of pages for batching, however,
4407 * causes the pages to actually be freed in smaller chunks. As there
4408 * can be a significant delay between the individual batches being
4409 * recycled, this leads to the once large chunks of space being
4410 * fragmented and becoming unavailable for high-order allocations.
4417 * pcp->high and pcp->batch values are related and dependent on one another:
4418 * ->batch must never be higher then ->high.
4419 * The following function updates them in a safe manner without read side
4422 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4423 * those fields changing asynchronously (acording the the above rule).
4425 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4426 * outside of boot time (or some other assurance that no concurrent updaters
4429 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4430 unsigned long batch)
4432 /* start with a fail safe value for batch */
4436 /* Update high, then batch, in order */
4443 /* a companion to pageset_set_high() */
4444 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4446 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4449 static void pageset_init(struct per_cpu_pageset *p)
4451 struct per_cpu_pages *pcp;
4454 memset(p, 0, sizeof(*p));
4458 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4459 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4462 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4465 pageset_set_batch(p, batch);
4469 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4470 * to the value high for the pageset p.
4472 static void pageset_set_high(struct per_cpu_pageset *p,
4475 unsigned long batch = max(1UL, high / 4);
4476 if ((high / 4) > (PAGE_SHIFT * 8))
4477 batch = PAGE_SHIFT * 8;
4479 pageset_update(&p->pcp, high, batch);
4482 static void pageset_set_high_and_batch(struct zone *zone,
4483 struct per_cpu_pageset *pcp)
4485 if (percpu_pagelist_fraction)
4486 pageset_set_high(pcp,
4487 (zone->managed_pages /
4488 percpu_pagelist_fraction));
4490 pageset_set_batch(pcp, zone_batchsize(zone));
4493 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4495 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4498 pageset_set_high_and_batch(zone, pcp);
4501 static void __meminit setup_zone_pageset(struct zone *zone)
4504 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4505 for_each_possible_cpu(cpu)
4506 zone_pageset_init(zone, cpu);
4510 * Allocate per cpu pagesets and initialize them.
4511 * Before this call only boot pagesets were available.
4513 void __init setup_per_cpu_pageset(void)
4517 for_each_populated_zone(zone)
4518 setup_zone_pageset(zone);
4521 static noinline __init_refok
4522 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4528 * The per-page waitqueue mechanism uses hashed waitqueues
4531 zone->wait_table_hash_nr_entries =
4532 wait_table_hash_nr_entries(zone_size_pages);
4533 zone->wait_table_bits =
4534 wait_table_bits(zone->wait_table_hash_nr_entries);
4535 alloc_size = zone->wait_table_hash_nr_entries
4536 * sizeof(wait_queue_head_t);
4538 if (!slab_is_available()) {
4539 zone->wait_table = (wait_queue_head_t *)
4540 memblock_virt_alloc_node_nopanic(
4541 alloc_size, zone->zone_pgdat->node_id);
4544 * This case means that a zone whose size was 0 gets new memory
4545 * via memory hot-add.
4546 * But it may be the case that a new node was hot-added. In
4547 * this case vmalloc() will not be able to use this new node's
4548 * memory - this wait_table must be initialized to use this new
4549 * node itself as well.
4550 * To use this new node's memory, further consideration will be
4553 zone->wait_table = vmalloc(alloc_size);
4555 if (!zone->wait_table)
4558 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4559 init_waitqueue_head(zone->wait_table + i);
4564 static __meminit void zone_pcp_init(struct zone *zone)
4567 * per cpu subsystem is not up at this point. The following code
4568 * relies on the ability of the linker to provide the
4569 * offset of a (static) per cpu variable into the per cpu area.
4571 zone->pageset = &boot_pageset;
4573 if (populated_zone(zone))
4574 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4575 zone->name, zone->present_pages,
4576 zone_batchsize(zone));
4579 int __meminit init_currently_empty_zone(struct zone *zone,
4580 unsigned long zone_start_pfn,
4582 enum memmap_context context)
4584 struct pglist_data *pgdat = zone->zone_pgdat;
4586 ret = zone_wait_table_init(zone, size);
4589 pgdat->nr_zones = zone_idx(zone) + 1;
4591 zone->zone_start_pfn = zone_start_pfn;
4593 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4594 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4596 (unsigned long)zone_idx(zone),
4597 zone_start_pfn, (zone_start_pfn + size));
4599 zone_init_free_lists(zone);
4604 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4605 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4608 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4610 int __meminit __early_pfn_to_nid(unsigned long pfn,
4611 struct mminit_pfnnid_cache *state)
4613 unsigned long start_pfn, end_pfn;
4616 if (state->last_start <= pfn && pfn < state->last_end)
4617 return state->last_nid;
4619 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4621 state->last_start = start_pfn;
4622 state->last_end = end_pfn;
4623 state->last_nid = nid;
4628 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4631 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4632 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4633 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4635 * If an architecture guarantees that all ranges registered contain no holes
4636 * and may be freed, this this function may be used instead of calling
4637 * memblock_free_early_nid() manually.
4639 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4641 unsigned long start_pfn, end_pfn;
4644 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4645 start_pfn = min(start_pfn, max_low_pfn);
4646 end_pfn = min(end_pfn, max_low_pfn);
4648 if (start_pfn < end_pfn)
4649 memblock_free_early_nid(PFN_PHYS(start_pfn),
4650 (end_pfn - start_pfn) << PAGE_SHIFT,
4656 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4657 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4659 * If an architecture guarantees that all ranges registered contain no holes and may
4660 * be freed, this function may be used instead of calling memory_present() manually.
4662 void __init sparse_memory_present_with_active_regions(int nid)
4664 unsigned long start_pfn, end_pfn;
4667 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4668 memory_present(this_nid, start_pfn, end_pfn);
4672 * get_pfn_range_for_nid - Return the start and end page frames for a node
4673 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4674 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4675 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4677 * It returns the start and end page frame of a node based on information
4678 * provided by memblock_set_node(). If called for a node
4679 * with no available memory, a warning is printed and the start and end
4682 void __meminit get_pfn_range_for_nid(unsigned int nid,
4683 unsigned long *start_pfn, unsigned long *end_pfn)
4685 unsigned long this_start_pfn, this_end_pfn;
4691 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4692 *start_pfn = min(*start_pfn, this_start_pfn);
4693 *end_pfn = max(*end_pfn, this_end_pfn);
4696 if (*start_pfn == -1UL)
4701 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4702 * assumption is made that zones within a node are ordered in monotonic
4703 * increasing memory addresses so that the "highest" populated zone is used
4705 static void __init find_usable_zone_for_movable(void)
4708 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4709 if (zone_index == ZONE_MOVABLE)
4712 if (arch_zone_highest_possible_pfn[zone_index] >
4713 arch_zone_lowest_possible_pfn[zone_index])
4717 VM_BUG_ON(zone_index == -1);
4718 movable_zone = zone_index;
4722 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4723 * because it is sized independent of architecture. Unlike the other zones,
4724 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4725 * in each node depending on the size of each node and how evenly kernelcore
4726 * is distributed. This helper function adjusts the zone ranges
4727 * provided by the architecture for a given node by using the end of the
4728 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4729 * zones within a node are in order of monotonic increases memory addresses
4731 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4732 unsigned long zone_type,
4733 unsigned long node_start_pfn,
4734 unsigned long node_end_pfn,
4735 unsigned long *zone_start_pfn,
4736 unsigned long *zone_end_pfn)
4738 /* Only adjust if ZONE_MOVABLE is on this node */
4739 if (zone_movable_pfn[nid]) {
4740 /* Size ZONE_MOVABLE */
4741 if (zone_type == ZONE_MOVABLE) {
4742 *zone_start_pfn = zone_movable_pfn[nid];
4743 *zone_end_pfn = min(node_end_pfn,
4744 arch_zone_highest_possible_pfn[movable_zone]);
4746 /* Adjust for ZONE_MOVABLE starting within this range */
4747 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4748 *zone_end_pfn > zone_movable_pfn[nid]) {
4749 *zone_end_pfn = zone_movable_pfn[nid];
4751 /* Check if this whole range is within ZONE_MOVABLE */
4752 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4753 *zone_start_pfn = *zone_end_pfn;
4758 * Return the number of pages a zone spans in a node, including holes
4759 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4761 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4762 unsigned long zone_type,
4763 unsigned long node_start_pfn,
4764 unsigned long node_end_pfn,
4765 unsigned long *ignored)
4767 unsigned long zone_start_pfn, zone_end_pfn;
4769 /* Get the start and end of the zone */
4770 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4771 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4772 adjust_zone_range_for_zone_movable(nid, zone_type,
4773 node_start_pfn, node_end_pfn,
4774 &zone_start_pfn, &zone_end_pfn);
4776 /* Check that this node has pages within the zone's required range */
4777 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4780 /* Move the zone boundaries inside the node if necessary */
4781 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4782 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4784 /* Return the spanned pages */
4785 return zone_end_pfn - zone_start_pfn;
4789 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4790 * then all holes in the requested range will be accounted for.
4792 unsigned long __meminit __absent_pages_in_range(int nid,
4793 unsigned long range_start_pfn,
4794 unsigned long range_end_pfn)
4796 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4797 unsigned long start_pfn, end_pfn;
4800 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4801 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4802 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4803 nr_absent -= end_pfn - start_pfn;
4809 * absent_pages_in_range - Return number of page frames in holes within a range
4810 * @start_pfn: The start PFN to start searching for holes
4811 * @end_pfn: The end PFN to stop searching for holes
4813 * It returns the number of pages frames in memory holes within a range.
4815 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4816 unsigned long end_pfn)
4818 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4821 /* Return the number of page frames in holes in a zone on a node */
4822 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4823 unsigned long zone_type,
4824 unsigned long node_start_pfn,
4825 unsigned long node_end_pfn,
4826 unsigned long *ignored)
4828 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4829 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4830 unsigned long zone_start_pfn, zone_end_pfn;
4832 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4833 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4835 adjust_zone_range_for_zone_movable(nid, zone_type,
4836 node_start_pfn, node_end_pfn,
4837 &zone_start_pfn, &zone_end_pfn);
4838 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4841 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4842 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4843 unsigned long zone_type,
4844 unsigned long node_start_pfn,
4845 unsigned long node_end_pfn,
4846 unsigned long *zones_size)
4848 return zones_size[zone_type];
4851 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4852 unsigned long zone_type,
4853 unsigned long node_start_pfn,
4854 unsigned long node_end_pfn,
4855 unsigned long *zholes_size)
4860 return zholes_size[zone_type];
4863 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4865 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4866 unsigned long node_start_pfn,
4867 unsigned long node_end_pfn,
4868 unsigned long *zones_size,
4869 unsigned long *zholes_size)
4871 unsigned long realtotalpages = 0, totalpages = 0;
4874 for (i = 0; i < MAX_NR_ZONES; i++) {
4875 struct zone *zone = pgdat->node_zones + i;
4876 unsigned long size, real_size;
4878 size = zone_spanned_pages_in_node(pgdat->node_id, i,
4882 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
4883 node_start_pfn, node_end_pfn,
4885 zone->spanned_pages = size;
4886 zone->present_pages = real_size;
4889 realtotalpages += real_size;
4892 pgdat->node_spanned_pages = totalpages;
4893 pgdat->node_present_pages = realtotalpages;
4894 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4898 #ifndef CONFIG_SPARSEMEM
4900 * Calculate the size of the zone->blockflags rounded to an unsigned long
4901 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4902 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4903 * round what is now in bits to nearest long in bits, then return it in
4906 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4908 unsigned long usemapsize;
4910 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4911 usemapsize = roundup(zonesize, pageblock_nr_pages);
4912 usemapsize = usemapsize >> pageblock_order;
4913 usemapsize *= NR_PAGEBLOCK_BITS;
4914 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4916 return usemapsize / 8;
4919 static void __init setup_usemap(struct pglist_data *pgdat,
4921 unsigned long zone_start_pfn,
4922 unsigned long zonesize)
4924 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4925 zone->pageblock_flags = NULL;
4927 zone->pageblock_flags =
4928 memblock_virt_alloc_node_nopanic(usemapsize,
4932 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4933 unsigned long zone_start_pfn, unsigned long zonesize) {}
4934 #endif /* CONFIG_SPARSEMEM */
4936 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4938 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4939 void __paginginit set_pageblock_order(void)
4943 /* Check that pageblock_nr_pages has not already been setup */
4944 if (pageblock_order)
4947 if (HPAGE_SHIFT > PAGE_SHIFT)
4948 order = HUGETLB_PAGE_ORDER;
4950 order = MAX_ORDER - 1;
4953 * Assume the largest contiguous order of interest is a huge page.
4954 * This value may be variable depending on boot parameters on IA64 and
4957 pageblock_order = order;
4959 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4962 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4963 * is unused as pageblock_order is set at compile-time. See
4964 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4967 void __paginginit set_pageblock_order(void)
4971 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4973 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4974 unsigned long present_pages)
4976 unsigned long pages = spanned_pages;
4979 * Provide a more accurate estimation if there are holes within
4980 * the zone and SPARSEMEM is in use. If there are holes within the
4981 * zone, each populated memory region may cost us one or two extra
4982 * memmap pages due to alignment because memmap pages for each
4983 * populated regions may not naturally algined on page boundary.
4984 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4986 if (spanned_pages > present_pages + (present_pages >> 4) &&
4987 IS_ENABLED(CONFIG_SPARSEMEM))
4988 pages = present_pages;
4990 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4994 * Set up the zone data structures:
4995 * - mark all pages reserved
4996 * - mark all memory queues empty
4997 * - clear the memory bitmaps
4999 * NOTE: pgdat should get zeroed by caller.
5001 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
5002 unsigned long node_start_pfn, unsigned long node_end_pfn)
5005 int nid = pgdat->node_id;
5006 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5009 pgdat_resize_init(pgdat);
5010 #ifdef CONFIG_NUMA_BALANCING
5011 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5012 pgdat->numabalancing_migrate_nr_pages = 0;
5013 pgdat->numabalancing_migrate_next_window = jiffies;
5015 init_waitqueue_head(&pgdat->kswapd_wait);
5016 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5017 pgdat_page_ext_init(pgdat);
5019 for (j = 0; j < MAX_NR_ZONES; j++) {
5020 struct zone *zone = pgdat->node_zones + j;
5021 unsigned long size, realsize, freesize, memmap_pages;
5023 size = zone->spanned_pages;
5024 realsize = freesize = zone->present_pages;
5027 * Adjust freesize so that it accounts for how much memory
5028 * is used by this zone for memmap. This affects the watermark
5029 * and per-cpu initialisations
5031 memmap_pages = calc_memmap_size(size, realsize);
5032 if (!is_highmem_idx(j)) {
5033 if (freesize >= memmap_pages) {
5034 freesize -= memmap_pages;
5037 " %s zone: %lu pages used for memmap\n",
5038 zone_names[j], memmap_pages);
5041 " %s zone: %lu pages exceeds freesize %lu\n",
5042 zone_names[j], memmap_pages, freesize);
5045 /* Account for reserved pages */
5046 if (j == 0 && freesize > dma_reserve) {
5047 freesize -= dma_reserve;
5048 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5049 zone_names[0], dma_reserve);
5052 if (!is_highmem_idx(j))
5053 nr_kernel_pages += freesize;
5054 /* Charge for highmem memmap if there are enough kernel pages */
5055 else if (nr_kernel_pages > memmap_pages * 2)
5056 nr_kernel_pages -= memmap_pages;
5057 nr_all_pages += freesize;
5060 * Set an approximate value for lowmem here, it will be adjusted
5061 * when the bootmem allocator frees pages into the buddy system.
5062 * And all highmem pages will be managed by the buddy system.
5064 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5067 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5069 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5071 zone->name = zone_names[j];
5072 spin_lock_init(&zone->lock);
5073 spin_lock_init(&zone->lru_lock);
5074 zone_seqlock_init(zone);
5075 zone->zone_pgdat = pgdat;
5076 zone_pcp_init(zone);
5078 /* For bootup, initialized properly in watermark setup */
5079 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5081 lruvec_init(&zone->lruvec);
5085 set_pageblock_order();
5086 setup_usemap(pgdat, zone, zone_start_pfn, size);
5087 ret = init_currently_empty_zone(zone, zone_start_pfn,
5088 size, MEMMAP_EARLY);
5090 memmap_init(size, nid, j, zone_start_pfn);
5091 zone_start_pfn += size;
5095 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5097 /* Skip empty nodes */
5098 if (!pgdat->node_spanned_pages)
5101 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5102 /* ia64 gets its own node_mem_map, before this, without bootmem */
5103 if (!pgdat->node_mem_map) {
5104 unsigned long size, start, end;
5108 * The zone's endpoints aren't required to be MAX_ORDER
5109 * aligned but the node_mem_map endpoints must be in order
5110 * for the buddy allocator to function correctly.
5112 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5113 end = pgdat_end_pfn(pgdat);
5114 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5115 size = (end - start) * sizeof(struct page);
5116 map = alloc_remap(pgdat->node_id, size);
5118 map = memblock_virt_alloc_node_nopanic(size,
5120 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5122 #ifndef CONFIG_NEED_MULTIPLE_NODES
5124 * With no DISCONTIG, the global mem_map is just set as node 0's
5126 if (pgdat == NODE_DATA(0)) {
5127 mem_map = NODE_DATA(0)->node_mem_map;
5128 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5129 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5130 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5131 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5134 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5137 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5138 unsigned long node_start_pfn, unsigned long *zholes_size)
5140 pg_data_t *pgdat = NODE_DATA(nid);
5141 unsigned long start_pfn = 0;
5142 unsigned long end_pfn = 0;
5144 /* pg_data_t should be reset to zero when it's allocated */
5145 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5147 pgdat->node_id = nid;
5148 pgdat->node_start_pfn = node_start_pfn;
5149 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5150 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5151 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5152 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5154 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5155 zones_size, zholes_size);
5157 alloc_node_mem_map(pgdat);
5158 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5159 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5160 nid, (unsigned long)pgdat,
5161 (unsigned long)pgdat->node_mem_map);
5164 free_area_init_core(pgdat, start_pfn, end_pfn);
5167 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5169 #if MAX_NUMNODES > 1
5171 * Figure out the number of possible node ids.
5173 void __init setup_nr_node_ids(void)
5176 unsigned int highest = 0;
5178 for_each_node_mask(node, node_possible_map)
5180 nr_node_ids = highest + 1;
5185 * node_map_pfn_alignment - determine the maximum internode alignment
5187 * This function should be called after node map is populated and sorted.
5188 * It calculates the maximum power of two alignment which can distinguish
5191 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5192 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5193 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5194 * shifted, 1GiB is enough and this function will indicate so.
5196 * This is used to test whether pfn -> nid mapping of the chosen memory
5197 * model has fine enough granularity to avoid incorrect mapping for the
5198 * populated node map.
5200 * Returns the determined alignment in pfn's. 0 if there is no alignment
5201 * requirement (single node).
5203 unsigned long __init node_map_pfn_alignment(void)
5205 unsigned long accl_mask = 0, last_end = 0;
5206 unsigned long start, end, mask;
5210 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5211 if (!start || last_nid < 0 || last_nid == nid) {
5218 * Start with a mask granular enough to pin-point to the
5219 * start pfn and tick off bits one-by-one until it becomes
5220 * too coarse to separate the current node from the last.
5222 mask = ~((1 << __ffs(start)) - 1);
5223 while (mask && last_end <= (start & (mask << 1)))
5226 /* accumulate all internode masks */
5230 /* convert mask to number of pages */
5231 return ~accl_mask + 1;
5234 /* Find the lowest pfn for a node */
5235 static unsigned long __init find_min_pfn_for_node(int nid)
5237 unsigned long min_pfn = ULONG_MAX;
5238 unsigned long start_pfn;
5241 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5242 min_pfn = min(min_pfn, start_pfn);
5244 if (min_pfn == ULONG_MAX) {
5246 "Could not find start_pfn for node %d\n", nid);
5254 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5256 * It returns the minimum PFN based on information provided via
5257 * memblock_set_node().
5259 unsigned long __init find_min_pfn_with_active_regions(void)
5261 return find_min_pfn_for_node(MAX_NUMNODES);
5265 * early_calculate_totalpages()
5266 * Sum pages in active regions for movable zone.
5267 * Populate N_MEMORY for calculating usable_nodes.
5269 static unsigned long __init early_calculate_totalpages(void)
5271 unsigned long totalpages = 0;
5272 unsigned long start_pfn, end_pfn;
5275 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5276 unsigned long pages = end_pfn - start_pfn;
5278 totalpages += pages;
5280 node_set_state(nid, N_MEMORY);
5286 * Find the PFN the Movable zone begins in each node. Kernel memory
5287 * is spread evenly between nodes as long as the nodes have enough
5288 * memory. When they don't, some nodes will have more kernelcore than
5291 static void __init find_zone_movable_pfns_for_nodes(void)
5294 unsigned long usable_startpfn;
5295 unsigned long kernelcore_node, kernelcore_remaining;
5296 /* save the state before borrow the nodemask */
5297 nodemask_t saved_node_state = node_states[N_MEMORY];
5298 unsigned long totalpages = early_calculate_totalpages();
5299 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5300 struct memblock_region *r;
5302 /* Need to find movable_zone earlier when movable_node is specified. */
5303 find_usable_zone_for_movable();
5306 * If movable_node is specified, ignore kernelcore and movablecore
5309 if (movable_node_is_enabled()) {
5310 for_each_memblock(memory, r) {
5311 if (!memblock_is_hotpluggable(r))
5316 usable_startpfn = PFN_DOWN(r->base);
5317 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5318 min(usable_startpfn, zone_movable_pfn[nid]) :
5326 * If movablecore=nn[KMG] was specified, calculate what size of
5327 * kernelcore that corresponds so that memory usable for
5328 * any allocation type is evenly spread. If both kernelcore
5329 * and movablecore are specified, then the value of kernelcore
5330 * will be used for required_kernelcore if it's greater than
5331 * what movablecore would have allowed.
5333 if (required_movablecore) {
5334 unsigned long corepages;
5337 * Round-up so that ZONE_MOVABLE is at least as large as what
5338 * was requested by the user
5340 required_movablecore =
5341 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5342 corepages = totalpages - required_movablecore;
5344 required_kernelcore = max(required_kernelcore, corepages);
5347 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5348 if (!required_kernelcore)
5351 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5352 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5355 /* Spread kernelcore memory as evenly as possible throughout nodes */
5356 kernelcore_node = required_kernelcore / usable_nodes;
5357 for_each_node_state(nid, N_MEMORY) {
5358 unsigned long start_pfn, end_pfn;
5361 * Recalculate kernelcore_node if the division per node
5362 * now exceeds what is necessary to satisfy the requested
5363 * amount of memory for the kernel
5365 if (required_kernelcore < kernelcore_node)
5366 kernelcore_node = required_kernelcore / usable_nodes;
5369 * As the map is walked, we track how much memory is usable
5370 * by the kernel using kernelcore_remaining. When it is
5371 * 0, the rest of the node is usable by ZONE_MOVABLE
5373 kernelcore_remaining = kernelcore_node;
5375 /* Go through each range of PFNs within this node */
5376 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5377 unsigned long size_pages;
5379 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5380 if (start_pfn >= end_pfn)
5383 /* Account for what is only usable for kernelcore */
5384 if (start_pfn < usable_startpfn) {
5385 unsigned long kernel_pages;
5386 kernel_pages = min(end_pfn, usable_startpfn)
5389 kernelcore_remaining -= min(kernel_pages,
5390 kernelcore_remaining);
5391 required_kernelcore -= min(kernel_pages,
5392 required_kernelcore);
5394 /* Continue if range is now fully accounted */
5395 if (end_pfn <= usable_startpfn) {
5398 * Push zone_movable_pfn to the end so
5399 * that if we have to rebalance
5400 * kernelcore across nodes, we will
5401 * not double account here
5403 zone_movable_pfn[nid] = end_pfn;
5406 start_pfn = usable_startpfn;
5410 * The usable PFN range for ZONE_MOVABLE is from
5411 * start_pfn->end_pfn. Calculate size_pages as the
5412 * number of pages used as kernelcore
5414 size_pages = end_pfn - start_pfn;
5415 if (size_pages > kernelcore_remaining)
5416 size_pages = kernelcore_remaining;
5417 zone_movable_pfn[nid] = start_pfn + size_pages;
5420 * Some kernelcore has been met, update counts and
5421 * break if the kernelcore for this node has been
5424 required_kernelcore -= min(required_kernelcore,
5426 kernelcore_remaining -= size_pages;
5427 if (!kernelcore_remaining)
5433 * If there is still required_kernelcore, we do another pass with one
5434 * less node in the count. This will push zone_movable_pfn[nid] further
5435 * along on the nodes that still have memory until kernelcore is
5439 if (usable_nodes && required_kernelcore > usable_nodes)
5443 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5444 for (nid = 0; nid < MAX_NUMNODES; nid++)
5445 zone_movable_pfn[nid] =
5446 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5449 /* restore the node_state */
5450 node_states[N_MEMORY] = saved_node_state;
5453 /* Any regular or high memory on that node ? */
5454 static void check_for_memory(pg_data_t *pgdat, int nid)
5456 enum zone_type zone_type;
5458 if (N_MEMORY == N_NORMAL_MEMORY)
5461 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5462 struct zone *zone = &pgdat->node_zones[zone_type];
5463 if (populated_zone(zone)) {
5464 node_set_state(nid, N_HIGH_MEMORY);
5465 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5466 zone_type <= ZONE_NORMAL)
5467 node_set_state(nid, N_NORMAL_MEMORY);
5474 * free_area_init_nodes - Initialise all pg_data_t and zone data
5475 * @max_zone_pfn: an array of max PFNs for each zone
5477 * This will call free_area_init_node() for each active node in the system.
5478 * Using the page ranges provided by memblock_set_node(), the size of each
5479 * zone in each node and their holes is calculated. If the maximum PFN
5480 * between two adjacent zones match, it is assumed that the zone is empty.
5481 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5482 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5483 * starts where the previous one ended. For example, ZONE_DMA32 starts
5484 * at arch_max_dma_pfn.
5486 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5488 unsigned long start_pfn, end_pfn;
5491 /* Record where the zone boundaries are */
5492 memset(arch_zone_lowest_possible_pfn, 0,
5493 sizeof(arch_zone_lowest_possible_pfn));
5494 memset(arch_zone_highest_possible_pfn, 0,
5495 sizeof(arch_zone_highest_possible_pfn));
5496 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5497 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5498 for (i = 1; i < MAX_NR_ZONES; i++) {
5499 if (i == ZONE_MOVABLE)
5501 arch_zone_lowest_possible_pfn[i] =
5502 arch_zone_highest_possible_pfn[i-1];
5503 arch_zone_highest_possible_pfn[i] =
5504 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5506 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5507 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5509 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5510 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5511 find_zone_movable_pfns_for_nodes();
5513 /* Print out the zone ranges */
5514 pr_info("Zone ranges:\n");
5515 for (i = 0; i < MAX_NR_ZONES; i++) {
5516 if (i == ZONE_MOVABLE)
5518 pr_info(" %-8s ", zone_names[i]);
5519 if (arch_zone_lowest_possible_pfn[i] ==
5520 arch_zone_highest_possible_pfn[i])
5523 pr_cont("[mem %#018Lx-%#018Lx]\n",
5524 (u64)arch_zone_lowest_possible_pfn[i]
5526 ((u64)arch_zone_highest_possible_pfn[i]
5527 << PAGE_SHIFT) - 1);
5530 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5531 pr_info("Movable zone start for each node\n");
5532 for (i = 0; i < MAX_NUMNODES; i++) {
5533 if (zone_movable_pfn[i])
5534 pr_info(" Node %d: %#018Lx\n", i,
5535 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5538 /* Print out the early node map */
5539 pr_info("Early memory node ranges\n");
5540 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5541 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5542 (u64)start_pfn << PAGE_SHIFT,
5543 ((u64)end_pfn << PAGE_SHIFT) - 1);
5545 /* Initialise every node */
5546 mminit_verify_pageflags_layout();
5547 setup_nr_node_ids();
5548 for_each_online_node(nid) {
5549 pg_data_t *pgdat = NODE_DATA(nid);
5550 free_area_init_node(nid, NULL,
5551 find_min_pfn_for_node(nid), NULL);
5553 /* Any memory on that node */
5554 if (pgdat->node_present_pages)
5555 node_set_state(nid, N_MEMORY);
5556 check_for_memory(pgdat, nid);
5560 static int __init cmdline_parse_core(char *p, unsigned long *core)
5562 unsigned long long coremem;
5566 coremem = memparse(p, &p);
5567 *core = coremem >> PAGE_SHIFT;
5569 /* Paranoid check that UL is enough for the coremem value */
5570 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5576 * kernelcore=size sets the amount of memory for use for allocations that
5577 * cannot be reclaimed or migrated.
5579 static int __init cmdline_parse_kernelcore(char *p)
5581 return cmdline_parse_core(p, &required_kernelcore);
5585 * movablecore=size sets the amount of memory for use for allocations that
5586 * can be reclaimed or migrated.
5588 static int __init cmdline_parse_movablecore(char *p)
5590 return cmdline_parse_core(p, &required_movablecore);
5593 early_param("kernelcore", cmdline_parse_kernelcore);
5594 early_param("movablecore", cmdline_parse_movablecore);
5596 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5598 void adjust_managed_page_count(struct page *page, long count)
5600 spin_lock(&managed_page_count_lock);
5601 page_zone(page)->managed_pages += count;
5602 totalram_pages += count;
5603 #ifdef CONFIG_HIGHMEM
5604 if (PageHighMem(page))
5605 totalhigh_pages += count;
5607 spin_unlock(&managed_page_count_lock);
5609 EXPORT_SYMBOL(adjust_managed_page_count);
5611 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5614 unsigned long pages = 0;
5616 start = (void *)PAGE_ALIGN((unsigned long)start);
5617 end = (void *)((unsigned long)end & PAGE_MASK);
5618 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5619 if ((unsigned int)poison <= 0xFF)
5620 memset(pos, poison, PAGE_SIZE);
5621 free_reserved_page(virt_to_page(pos));
5625 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5626 s, pages << (PAGE_SHIFT - 10), start, end);
5630 EXPORT_SYMBOL(free_reserved_area);
5632 #ifdef CONFIG_HIGHMEM
5633 void free_highmem_page(struct page *page)
5635 __free_reserved_page(page);
5637 page_zone(page)->managed_pages++;
5643 void __init mem_init_print_info(const char *str)
5645 unsigned long physpages, codesize, datasize, rosize, bss_size;
5646 unsigned long init_code_size, init_data_size;
5648 physpages = get_num_physpages();
5649 codesize = _etext - _stext;
5650 datasize = _edata - _sdata;
5651 rosize = __end_rodata - __start_rodata;
5652 bss_size = __bss_stop - __bss_start;
5653 init_data_size = __init_end - __init_begin;
5654 init_code_size = _einittext - _sinittext;
5657 * Detect special cases and adjust section sizes accordingly:
5658 * 1) .init.* may be embedded into .data sections
5659 * 2) .init.text.* may be out of [__init_begin, __init_end],
5660 * please refer to arch/tile/kernel/vmlinux.lds.S.
5661 * 3) .rodata.* may be embedded into .text or .data sections.
5663 #define adj_init_size(start, end, size, pos, adj) \
5665 if (start <= pos && pos < end && size > adj) \
5669 adj_init_size(__init_begin, __init_end, init_data_size,
5670 _sinittext, init_code_size);
5671 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5672 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5673 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5674 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5676 #undef adj_init_size
5678 pr_info("Memory: %luK/%luK available "
5679 "(%luK kernel code, %luK rwdata, %luK rodata, "
5680 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5681 #ifdef CONFIG_HIGHMEM
5685 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5686 codesize >> 10, datasize >> 10, rosize >> 10,
5687 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5688 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5689 totalcma_pages << (PAGE_SHIFT-10),
5690 #ifdef CONFIG_HIGHMEM
5691 totalhigh_pages << (PAGE_SHIFT-10),
5693 str ? ", " : "", str ? str : "");
5697 * set_dma_reserve - set the specified number of pages reserved in the first zone
5698 * @new_dma_reserve: The number of pages to mark reserved
5700 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5701 * In the DMA zone, a significant percentage may be consumed by kernel image
5702 * and other unfreeable allocations which can skew the watermarks badly. This
5703 * function may optionally be used to account for unfreeable pages in the
5704 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5705 * smaller per-cpu batchsize.
5707 void __init set_dma_reserve(unsigned long new_dma_reserve)
5709 dma_reserve = new_dma_reserve;
5712 void __init free_area_init(unsigned long *zones_size)
5714 free_area_init_node(0, zones_size,
5715 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5718 static int page_alloc_cpu_notify(struct notifier_block *self,
5719 unsigned long action, void *hcpu)
5721 int cpu = (unsigned long)hcpu;
5723 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5724 lru_add_drain_cpu(cpu);
5728 * Spill the event counters of the dead processor
5729 * into the current processors event counters.
5730 * This artificially elevates the count of the current
5733 vm_events_fold_cpu(cpu);
5736 * Zero the differential counters of the dead processor
5737 * so that the vm statistics are consistent.
5739 * This is only okay since the processor is dead and cannot
5740 * race with what we are doing.
5742 cpu_vm_stats_fold(cpu);
5747 void __init page_alloc_init(void)
5749 hotcpu_notifier(page_alloc_cpu_notify, 0);
5753 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5754 * or min_free_kbytes changes.
5756 static void calculate_totalreserve_pages(void)
5758 struct pglist_data *pgdat;
5759 unsigned long reserve_pages = 0;
5760 enum zone_type i, j;
5762 for_each_online_pgdat(pgdat) {
5763 for (i = 0; i < MAX_NR_ZONES; i++) {
5764 struct zone *zone = pgdat->node_zones + i;
5767 /* Find valid and maximum lowmem_reserve in the zone */
5768 for (j = i; j < MAX_NR_ZONES; j++) {
5769 if (zone->lowmem_reserve[j] > max)
5770 max = zone->lowmem_reserve[j];
5773 /* we treat the high watermark as reserved pages. */
5774 max += high_wmark_pages(zone);
5776 if (max > zone->managed_pages)
5777 max = zone->managed_pages;
5778 reserve_pages += max;
5780 * Lowmem reserves are not available to
5781 * GFP_HIGHUSER page cache allocations and
5782 * kswapd tries to balance zones to their high
5783 * watermark. As a result, neither should be
5784 * regarded as dirtyable memory, to prevent a
5785 * situation where reclaim has to clean pages
5786 * in order to balance the zones.
5788 zone->dirty_balance_reserve = max;
5791 dirty_balance_reserve = reserve_pages;
5792 totalreserve_pages = reserve_pages;
5796 * setup_per_zone_lowmem_reserve - called whenever
5797 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5798 * has a correct pages reserved value, so an adequate number of
5799 * pages are left in the zone after a successful __alloc_pages().
5801 static void setup_per_zone_lowmem_reserve(void)
5803 struct pglist_data *pgdat;
5804 enum zone_type j, idx;
5806 for_each_online_pgdat(pgdat) {
5807 for (j = 0; j < MAX_NR_ZONES; j++) {
5808 struct zone *zone = pgdat->node_zones + j;
5809 unsigned long managed_pages = zone->managed_pages;
5811 zone->lowmem_reserve[j] = 0;
5815 struct zone *lower_zone;
5819 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5820 sysctl_lowmem_reserve_ratio[idx] = 1;
5822 lower_zone = pgdat->node_zones + idx;
5823 lower_zone->lowmem_reserve[j] = managed_pages /
5824 sysctl_lowmem_reserve_ratio[idx];
5825 managed_pages += lower_zone->managed_pages;
5830 /* update totalreserve_pages */
5831 calculate_totalreserve_pages();
5834 static void __setup_per_zone_wmarks(void)
5836 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5837 unsigned long lowmem_pages = 0;
5839 unsigned long flags;
5841 /* Calculate total number of !ZONE_HIGHMEM pages */
5842 for_each_zone(zone) {
5843 if (!is_highmem(zone))
5844 lowmem_pages += zone->managed_pages;
5847 for_each_zone(zone) {
5850 spin_lock_irqsave(&zone->lock, flags);
5851 tmp = (u64)pages_min * zone->managed_pages;
5852 do_div(tmp, lowmem_pages);
5853 if (is_highmem(zone)) {
5855 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5856 * need highmem pages, so cap pages_min to a small
5859 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5860 * deltas control asynch page reclaim, and so should
5861 * not be capped for highmem.
5863 unsigned long min_pages;
5865 min_pages = zone->managed_pages / 1024;
5866 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5867 zone->watermark[WMARK_MIN] = min_pages;
5870 * If it's a lowmem zone, reserve a number of pages
5871 * proportionate to the zone's size.
5873 zone->watermark[WMARK_MIN] = tmp;
5876 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5877 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5879 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5880 high_wmark_pages(zone) - low_wmark_pages(zone) -
5881 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5883 setup_zone_migrate_reserve(zone);
5884 spin_unlock_irqrestore(&zone->lock, flags);
5887 /* update totalreserve_pages */
5888 calculate_totalreserve_pages();
5892 * setup_per_zone_wmarks - called when min_free_kbytes changes
5893 * or when memory is hot-{added|removed}
5895 * Ensures that the watermark[min,low,high] values for each zone are set
5896 * correctly with respect to min_free_kbytes.
5898 void setup_per_zone_wmarks(void)
5900 mutex_lock(&zonelists_mutex);
5901 __setup_per_zone_wmarks();
5902 mutex_unlock(&zonelists_mutex);
5906 * The inactive anon list should be small enough that the VM never has to
5907 * do too much work, but large enough that each inactive page has a chance
5908 * to be referenced again before it is swapped out.
5910 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5911 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5912 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5913 * the anonymous pages are kept on the inactive list.
5916 * memory ratio inactive anon
5917 * -------------------------------------
5926 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5928 unsigned int gb, ratio;
5930 /* Zone size in gigabytes */
5931 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5933 ratio = int_sqrt(10 * gb);
5937 zone->inactive_ratio = ratio;
5940 static void __meminit setup_per_zone_inactive_ratio(void)
5945 calculate_zone_inactive_ratio(zone);
5949 * Initialise min_free_kbytes.
5951 * For small machines we want it small (128k min). For large machines
5952 * we want it large (64MB max). But it is not linear, because network
5953 * bandwidth does not increase linearly with machine size. We use
5955 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5956 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5972 int __meminit init_per_zone_wmark_min(void)
5974 unsigned long lowmem_kbytes;
5975 int new_min_free_kbytes;
5977 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5978 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5980 if (new_min_free_kbytes > user_min_free_kbytes) {
5981 min_free_kbytes = new_min_free_kbytes;
5982 if (min_free_kbytes < 128)
5983 min_free_kbytes = 128;
5984 if (min_free_kbytes > 65536)
5985 min_free_kbytes = 65536;
5987 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5988 new_min_free_kbytes, user_min_free_kbytes);
5990 setup_per_zone_wmarks();
5991 refresh_zone_stat_thresholds();
5992 setup_per_zone_lowmem_reserve();
5993 setup_per_zone_inactive_ratio();
5996 module_init(init_per_zone_wmark_min)
5999 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6000 * that we can call two helper functions whenever min_free_kbytes
6003 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6004 void __user *buffer, size_t *length, loff_t *ppos)
6008 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6013 user_min_free_kbytes = min_free_kbytes;
6014 setup_per_zone_wmarks();
6020 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6021 void __user *buffer, size_t *length, loff_t *ppos)
6026 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6031 zone->min_unmapped_pages = (zone->managed_pages *
6032 sysctl_min_unmapped_ratio) / 100;
6036 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6037 void __user *buffer, size_t *length, loff_t *ppos)
6042 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6047 zone->min_slab_pages = (zone->managed_pages *
6048 sysctl_min_slab_ratio) / 100;
6054 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6055 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6056 * whenever sysctl_lowmem_reserve_ratio changes.
6058 * The reserve ratio obviously has absolutely no relation with the
6059 * minimum watermarks. The lowmem reserve ratio can only make sense
6060 * if in function of the boot time zone sizes.
6062 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6063 void __user *buffer, size_t *length, loff_t *ppos)
6065 proc_dointvec_minmax(table, write, buffer, length, ppos);
6066 setup_per_zone_lowmem_reserve();
6071 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6072 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6073 * pagelist can have before it gets flushed back to buddy allocator.
6075 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6076 void __user *buffer, size_t *length, loff_t *ppos)
6079 int old_percpu_pagelist_fraction;
6082 mutex_lock(&pcp_batch_high_lock);
6083 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6085 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6086 if (!write || ret < 0)
6089 /* Sanity checking to avoid pcp imbalance */
6090 if (percpu_pagelist_fraction &&
6091 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6092 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6098 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6101 for_each_populated_zone(zone) {
6104 for_each_possible_cpu(cpu)
6105 pageset_set_high_and_batch(zone,
6106 per_cpu_ptr(zone->pageset, cpu));
6109 mutex_unlock(&pcp_batch_high_lock);
6114 int hashdist = HASHDIST_DEFAULT;
6116 static int __init set_hashdist(char *str)
6120 hashdist = simple_strtoul(str, &str, 0);
6123 __setup("hashdist=", set_hashdist);
6127 * allocate a large system hash table from bootmem
6128 * - it is assumed that the hash table must contain an exact power-of-2
6129 * quantity of entries
6130 * - limit is the number of hash buckets, not the total allocation size
6132 void *__init alloc_large_system_hash(const char *tablename,
6133 unsigned long bucketsize,
6134 unsigned long numentries,
6137 unsigned int *_hash_shift,
6138 unsigned int *_hash_mask,
6139 unsigned long low_limit,
6140 unsigned long high_limit)
6142 unsigned long long max = high_limit;
6143 unsigned long log2qty, size;
6146 /* allow the kernel cmdline to have a say */
6148 /* round applicable memory size up to nearest megabyte */
6149 numentries = nr_kernel_pages;
6151 /* It isn't necessary when PAGE_SIZE >= 1MB */
6152 if (PAGE_SHIFT < 20)
6153 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6155 /* limit to 1 bucket per 2^scale bytes of low memory */
6156 if (scale > PAGE_SHIFT)
6157 numentries >>= (scale - PAGE_SHIFT);
6159 numentries <<= (PAGE_SHIFT - scale);
6161 /* Make sure we've got at least a 0-order allocation.. */
6162 if (unlikely(flags & HASH_SMALL)) {
6163 /* Makes no sense without HASH_EARLY */
6164 WARN_ON(!(flags & HASH_EARLY));
6165 if (!(numentries >> *_hash_shift)) {
6166 numentries = 1UL << *_hash_shift;
6167 BUG_ON(!numentries);
6169 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6170 numentries = PAGE_SIZE / bucketsize;
6172 numentries = roundup_pow_of_two(numentries);
6174 /* limit allocation size to 1/16 total memory by default */
6176 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6177 do_div(max, bucketsize);
6179 max = min(max, 0x80000000ULL);
6181 if (numentries < low_limit)
6182 numentries = low_limit;
6183 if (numentries > max)
6186 log2qty = ilog2(numentries);
6189 size = bucketsize << log2qty;
6190 if (flags & HASH_EARLY)
6191 table = memblock_virt_alloc_nopanic(size, 0);
6193 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6196 * If bucketsize is not a power-of-two, we may free
6197 * some pages at the end of hash table which
6198 * alloc_pages_exact() automatically does
6200 if (get_order(size) < MAX_ORDER) {
6201 table = alloc_pages_exact(size, GFP_ATOMIC);
6202 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6205 } while (!table && size > PAGE_SIZE && --log2qty);
6208 panic("Failed to allocate %s hash table\n", tablename);
6210 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6213 ilog2(size) - PAGE_SHIFT,
6217 *_hash_shift = log2qty;
6219 *_hash_mask = (1 << log2qty) - 1;
6224 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6225 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6228 #ifdef CONFIG_SPARSEMEM
6229 return __pfn_to_section(pfn)->pageblock_flags;
6231 return zone->pageblock_flags;
6232 #endif /* CONFIG_SPARSEMEM */
6235 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6237 #ifdef CONFIG_SPARSEMEM
6238 pfn &= (PAGES_PER_SECTION-1);
6239 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6241 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6242 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6243 #endif /* CONFIG_SPARSEMEM */
6247 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6248 * @page: The page within the block of interest
6249 * @pfn: The target page frame number
6250 * @end_bitidx: The last bit of interest to retrieve
6251 * @mask: mask of bits that the caller is interested in
6253 * Return: pageblock_bits flags
6255 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6256 unsigned long end_bitidx,
6260 unsigned long *bitmap;
6261 unsigned long bitidx, word_bitidx;
6264 zone = page_zone(page);
6265 bitmap = get_pageblock_bitmap(zone, pfn);
6266 bitidx = pfn_to_bitidx(zone, pfn);
6267 word_bitidx = bitidx / BITS_PER_LONG;
6268 bitidx &= (BITS_PER_LONG-1);
6270 word = bitmap[word_bitidx];
6271 bitidx += end_bitidx;
6272 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6276 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6277 * @page: The page within the block of interest
6278 * @flags: The flags to set
6279 * @pfn: The target page frame number
6280 * @end_bitidx: The last bit of interest
6281 * @mask: mask of bits that the caller is interested in
6283 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6285 unsigned long end_bitidx,
6289 unsigned long *bitmap;
6290 unsigned long bitidx, word_bitidx;
6291 unsigned long old_word, word;
6293 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6295 zone = page_zone(page);
6296 bitmap = get_pageblock_bitmap(zone, pfn);
6297 bitidx = pfn_to_bitidx(zone, pfn);
6298 word_bitidx = bitidx / BITS_PER_LONG;
6299 bitidx &= (BITS_PER_LONG-1);
6301 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6303 bitidx += end_bitidx;
6304 mask <<= (BITS_PER_LONG - bitidx - 1);
6305 flags <<= (BITS_PER_LONG - bitidx - 1);
6307 word = READ_ONCE(bitmap[word_bitidx]);
6309 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6310 if (word == old_word)
6317 * This function checks whether pageblock includes unmovable pages or not.
6318 * If @count is not zero, it is okay to include less @count unmovable pages
6320 * PageLRU check without isolation or lru_lock could race so that
6321 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6322 * expect this function should be exact.
6324 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6325 bool skip_hwpoisoned_pages)
6327 unsigned long pfn, iter, found;
6331 * For avoiding noise data, lru_add_drain_all() should be called
6332 * If ZONE_MOVABLE, the zone never contains unmovable pages
6334 if (zone_idx(zone) == ZONE_MOVABLE)
6336 mt = get_pageblock_migratetype(page);
6337 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6340 pfn = page_to_pfn(page);
6341 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6342 unsigned long check = pfn + iter;
6344 if (!pfn_valid_within(check))
6347 page = pfn_to_page(check);
6350 * Hugepages are not in LRU lists, but they're movable.
6351 * We need not scan over tail pages bacause we don't
6352 * handle each tail page individually in migration.
6354 if (PageHuge(page)) {
6355 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6360 * We can't use page_count without pin a page
6361 * because another CPU can free compound page.
6362 * This check already skips compound tails of THP
6363 * because their page->_count is zero at all time.
6365 if (!atomic_read(&page->_count)) {
6366 if (PageBuddy(page))
6367 iter += (1 << page_order(page)) - 1;
6372 * The HWPoisoned page may be not in buddy system, and
6373 * page_count() is not 0.
6375 if (skip_hwpoisoned_pages && PageHWPoison(page))
6381 * If there are RECLAIMABLE pages, we need to check
6382 * it. But now, memory offline itself doesn't call
6383 * shrink_node_slabs() and it still to be fixed.
6386 * If the page is not RAM, page_count()should be 0.
6387 * we don't need more check. This is an _used_ not-movable page.
6389 * The problematic thing here is PG_reserved pages. PG_reserved
6390 * is set to both of a memory hole page and a _used_ kernel
6399 bool is_pageblock_removable_nolock(struct page *page)
6405 * We have to be careful here because we are iterating over memory
6406 * sections which are not zone aware so we might end up outside of
6407 * the zone but still within the section.
6408 * We have to take care about the node as well. If the node is offline
6409 * its NODE_DATA will be NULL - see page_zone.
6411 if (!node_online(page_to_nid(page)))
6414 zone = page_zone(page);
6415 pfn = page_to_pfn(page);
6416 if (!zone_spans_pfn(zone, pfn))
6419 return !has_unmovable_pages(zone, page, 0, true);
6424 static unsigned long pfn_max_align_down(unsigned long pfn)
6426 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6427 pageblock_nr_pages) - 1);
6430 static unsigned long pfn_max_align_up(unsigned long pfn)
6432 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6433 pageblock_nr_pages));
6436 /* [start, end) must belong to a single zone. */
6437 static int __alloc_contig_migrate_range(struct compact_control *cc,
6438 unsigned long start, unsigned long end)
6440 /* This function is based on compact_zone() from compaction.c. */
6441 unsigned long nr_reclaimed;
6442 unsigned long pfn = start;
6443 unsigned int tries = 0;
6448 while (pfn < end || !list_empty(&cc->migratepages)) {
6449 if (fatal_signal_pending(current)) {
6454 if (list_empty(&cc->migratepages)) {
6455 cc->nr_migratepages = 0;
6456 pfn = isolate_migratepages_range(cc, pfn, end);
6462 } else if (++tries == 5) {
6463 ret = ret < 0 ? ret : -EBUSY;
6467 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6469 cc->nr_migratepages -= nr_reclaimed;
6471 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6472 NULL, 0, cc->mode, MR_CMA);
6475 putback_movable_pages(&cc->migratepages);
6482 * alloc_contig_range() -- tries to allocate given range of pages
6483 * @start: start PFN to allocate
6484 * @end: one-past-the-last PFN to allocate
6485 * @migratetype: migratetype of the underlaying pageblocks (either
6486 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6487 * in range must have the same migratetype and it must
6488 * be either of the two.
6490 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6491 * aligned, however it's the caller's responsibility to guarantee that
6492 * we are the only thread that changes migrate type of pageblocks the
6495 * The PFN range must belong to a single zone.
6497 * Returns zero on success or negative error code. On success all
6498 * pages which PFN is in [start, end) are allocated for the caller and
6499 * need to be freed with free_contig_range().
6501 int alloc_contig_range(unsigned long start, unsigned long end,
6502 unsigned migratetype)
6504 unsigned long outer_start, outer_end;
6507 struct compact_control cc = {
6508 .nr_migratepages = 0,
6510 .zone = page_zone(pfn_to_page(start)),
6511 .mode = MIGRATE_SYNC,
6512 .ignore_skip_hint = true,
6514 INIT_LIST_HEAD(&cc.migratepages);
6517 * What we do here is we mark all pageblocks in range as
6518 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6519 * have different sizes, and due to the way page allocator
6520 * work, we align the range to biggest of the two pages so
6521 * that page allocator won't try to merge buddies from
6522 * different pageblocks and change MIGRATE_ISOLATE to some
6523 * other migration type.
6525 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6526 * migrate the pages from an unaligned range (ie. pages that
6527 * we are interested in). This will put all the pages in
6528 * range back to page allocator as MIGRATE_ISOLATE.
6530 * When this is done, we take the pages in range from page
6531 * allocator removing them from the buddy system. This way
6532 * page allocator will never consider using them.
6534 * This lets us mark the pageblocks back as
6535 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6536 * aligned range but not in the unaligned, original range are
6537 * put back to page allocator so that buddy can use them.
6540 ret = start_isolate_page_range(pfn_max_align_down(start),
6541 pfn_max_align_up(end), migratetype,
6546 ret = __alloc_contig_migrate_range(&cc, start, end);
6551 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6552 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6553 * more, all pages in [start, end) are free in page allocator.
6554 * What we are going to do is to allocate all pages from
6555 * [start, end) (that is remove them from page allocator).
6557 * The only problem is that pages at the beginning and at the
6558 * end of interesting range may be not aligned with pages that
6559 * page allocator holds, ie. they can be part of higher order
6560 * pages. Because of this, we reserve the bigger range and
6561 * once this is done free the pages we are not interested in.
6563 * We don't have to hold zone->lock here because the pages are
6564 * isolated thus they won't get removed from buddy.
6567 lru_add_drain_all();
6568 drain_all_pages(cc.zone);
6571 outer_start = start;
6572 while (!PageBuddy(pfn_to_page(outer_start))) {
6573 if (++order >= MAX_ORDER) {
6577 outer_start &= ~0UL << order;
6580 /* Make sure the range is really isolated. */
6581 if (test_pages_isolated(outer_start, end, false)) {
6582 pr_info("%s: [%lx, %lx) PFNs busy\n",
6583 __func__, outer_start, end);
6588 /* Grab isolated pages from freelists. */
6589 outer_end = isolate_freepages_range(&cc, outer_start, end);
6595 /* Free head and tail (if any) */
6596 if (start != outer_start)
6597 free_contig_range(outer_start, start - outer_start);
6598 if (end != outer_end)
6599 free_contig_range(end, outer_end - end);
6602 undo_isolate_page_range(pfn_max_align_down(start),
6603 pfn_max_align_up(end), migratetype);
6607 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6609 unsigned int count = 0;
6611 for (; nr_pages--; pfn++) {
6612 struct page *page = pfn_to_page(pfn);
6614 count += page_count(page) != 1;
6617 WARN(count != 0, "%d pages are still in use!\n", count);
6621 #ifdef CONFIG_MEMORY_HOTPLUG
6623 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6624 * page high values need to be recalulated.
6626 void __meminit zone_pcp_update(struct zone *zone)
6629 mutex_lock(&pcp_batch_high_lock);
6630 for_each_possible_cpu(cpu)
6631 pageset_set_high_and_batch(zone,
6632 per_cpu_ptr(zone->pageset, cpu));
6633 mutex_unlock(&pcp_batch_high_lock);
6637 void zone_pcp_reset(struct zone *zone)
6639 unsigned long flags;
6641 struct per_cpu_pageset *pset;
6643 /* avoid races with drain_pages() */
6644 local_irq_save(flags);
6645 if (zone->pageset != &boot_pageset) {
6646 for_each_online_cpu(cpu) {
6647 pset = per_cpu_ptr(zone->pageset, cpu);
6648 drain_zonestat(zone, pset);
6650 free_percpu(zone->pageset);
6651 zone->pageset = &boot_pageset;
6653 local_irq_restore(flags);
6656 #ifdef CONFIG_MEMORY_HOTREMOVE
6658 * All pages in the range must be isolated before calling this.
6661 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6665 unsigned int order, i;
6667 unsigned long flags;
6668 /* find the first valid pfn */
6669 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6674 zone = page_zone(pfn_to_page(pfn));
6675 spin_lock_irqsave(&zone->lock, flags);
6677 while (pfn < end_pfn) {
6678 if (!pfn_valid(pfn)) {
6682 page = pfn_to_page(pfn);
6684 * The HWPoisoned page may be not in buddy system, and
6685 * page_count() is not 0.
6687 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6689 SetPageReserved(page);
6693 BUG_ON(page_count(page));
6694 BUG_ON(!PageBuddy(page));
6695 order = page_order(page);
6696 #ifdef CONFIG_DEBUG_VM
6697 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6698 pfn, 1 << order, end_pfn);
6700 list_del(&page->lru);
6701 rmv_page_order(page);
6702 zone->free_area[order].nr_free--;
6703 for (i = 0; i < (1 << order); i++)
6704 SetPageReserved((page+i));
6705 pfn += (1 << order);
6707 spin_unlock_irqrestore(&zone->lock, flags);
6711 #ifdef CONFIG_MEMORY_FAILURE
6712 bool is_free_buddy_page(struct page *page)
6714 struct zone *zone = page_zone(page);
6715 unsigned long pfn = page_to_pfn(page);
6716 unsigned long flags;
6719 spin_lock_irqsave(&zone->lock, flags);
6720 for (order = 0; order < MAX_ORDER; order++) {
6721 struct page *page_head = page - (pfn & ((1 << order) - 1));
6723 if (PageBuddy(page_head) && page_order(page_head) >= order)
6726 spin_unlock_irqrestore(&zone->lock, flags);
6728 return order < MAX_ORDER;