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/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node);
74 EXPORT_PER_CPU_SYMBOL(numa_node);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
86 int _node_numa_mem_[MAX_NUMNODES];
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
103 [N_CPU] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
208 * Try to keep at least this much lowmem free. Do not allow normal
209 * allocations below this point, only high priority ones. Automatically
210 * tuned according to the amount of memory in the system.
212 int min_free_kbytes = 1024;
213 int user_min_free_kbytes = -1;
216 * Extra memory for the system to try freeing. Used to temporarily
217 * free memory, to make space for new workloads. Anyone can allocate
218 * down to the min watermarks controlled by min_free_kbytes above.
220 int extra_free_kbytes = 0;
222 static unsigned long __meminitdata nr_kernel_pages;
223 static unsigned long __meminitdata nr_all_pages;
224 static unsigned long __meminitdata dma_reserve;
226 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
227 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
228 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
229 static unsigned long __initdata required_kernelcore;
230 static unsigned long __initdata required_movablecore;
231 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
233 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
235 EXPORT_SYMBOL(movable_zone);
236 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
239 int nr_node_ids __read_mostly = MAX_NUMNODES;
240 int nr_online_nodes __read_mostly = 1;
241 EXPORT_SYMBOL(nr_node_ids);
242 EXPORT_SYMBOL(nr_online_nodes);
245 int page_group_by_mobility_disabled __read_mostly;
247 void set_pageblock_migratetype(struct page *page, int migratetype)
249 if (unlikely(page_group_by_mobility_disabled &&
250 migratetype < MIGRATE_PCPTYPES))
251 migratetype = MIGRATE_UNMOVABLE;
253 set_pageblock_flags_group(page, (unsigned long)migratetype,
254 PB_migrate, PB_migrate_end);
257 bool oom_killer_disabled __read_mostly;
259 #ifdef CONFIG_DEBUG_VM
260 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
264 unsigned long pfn = page_to_pfn(page);
265 unsigned long sp, start_pfn;
268 seq = zone_span_seqbegin(zone);
269 start_pfn = zone->zone_start_pfn;
270 sp = zone->spanned_pages;
271 if (!zone_spans_pfn(zone, pfn))
273 } while (zone_span_seqretry(zone, seq));
276 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
277 pfn, zone_to_nid(zone), zone->name,
278 start_pfn, start_pfn + sp);
283 static int page_is_consistent(struct zone *zone, struct page *page)
285 if (!pfn_valid_within(page_to_pfn(page)))
287 if (zone != page_zone(page))
293 * Temporary debugging check for pages not lying within a given zone.
295 static int bad_range(struct zone *zone, struct page *page)
297 if (page_outside_zone_boundaries(zone, page))
299 if (!page_is_consistent(zone, page))
305 static inline int bad_range(struct zone *zone, struct page *page)
311 static void bad_page(struct page *page, const char *reason,
312 unsigned long bad_flags)
314 static unsigned long resume;
315 static unsigned long nr_shown;
316 static unsigned long nr_unshown;
318 /* Don't complain about poisoned pages */
319 if (PageHWPoison(page)) {
320 page_mapcount_reset(page); /* remove PageBuddy */
325 * Allow a burst of 60 reports, then keep quiet for that minute;
326 * or allow a steady drip of one report per second.
328 if (nr_shown == 60) {
329 if (time_before(jiffies, resume)) {
335 "BUG: Bad page state: %lu messages suppressed\n",
342 resume = jiffies + 60 * HZ;
344 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
345 current->comm, page_to_pfn(page));
346 dump_page_badflags(page, reason, bad_flags);
351 /* Leave bad fields for debug, except PageBuddy could make trouble */
352 page_mapcount_reset(page); /* remove PageBuddy */
353 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
357 * Higher-order pages are called "compound pages". They are structured thusly:
359 * The first PAGE_SIZE page is called the "head page".
361 * The remaining PAGE_SIZE pages are called "tail pages".
363 * All pages have PG_compound set. All tail pages have their ->first_page
364 * pointing at the head page.
366 * The first tail page's ->lru.next holds the address of the compound page's
367 * put_page() function. Its ->lru.prev holds the order of allocation.
368 * This usage means that zero-order pages may not be compound.
371 static void free_compound_page(struct page *page)
373 __free_pages_ok(page, compound_order(page));
376 void prep_compound_page(struct page *page, unsigned long order)
379 int nr_pages = 1 << order;
381 set_compound_page_dtor(page, free_compound_page);
382 set_compound_order(page, order);
384 for (i = 1; i < nr_pages; i++) {
385 struct page *p = page + i;
386 set_page_count(p, 0);
387 p->first_page = page;
388 /* Make sure p->first_page is always valid for PageTail() */
394 /* update __split_huge_page_refcount if you change this function */
395 static int destroy_compound_page(struct page *page, unsigned long order)
398 int nr_pages = 1 << order;
401 if (unlikely(compound_order(page) != order)) {
402 bad_page(page, "wrong compound order", 0);
406 __ClearPageHead(page);
408 for (i = 1; i < nr_pages; i++) {
409 struct page *p = page + i;
411 if (unlikely(!PageTail(p))) {
412 bad_page(page, "PageTail not set", 0);
414 } else if (unlikely(p->first_page != page)) {
415 bad_page(page, "first_page not consistent", 0);
424 static inline void prep_zero_page(struct page *page, unsigned int order,
430 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
431 * and __GFP_HIGHMEM from hard or soft interrupt context.
433 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
434 for (i = 0; i < (1 << order); i++)
435 clear_highpage(page + i);
438 #ifdef CONFIG_DEBUG_PAGEALLOC
439 unsigned int _debug_guardpage_minorder;
441 static int __init debug_guardpage_minorder_setup(char *buf)
445 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
446 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
449 _debug_guardpage_minorder = res;
450 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
453 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
455 static inline void set_page_guard_flag(struct page *page)
457 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
460 static inline void clear_page_guard_flag(struct page *page)
462 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
465 static inline void set_page_guard_flag(struct page *page) { }
466 static inline void clear_page_guard_flag(struct page *page) { }
469 static inline void set_page_order(struct page *page, unsigned int order)
471 set_page_private(page, order);
472 __SetPageBuddy(page);
475 static inline void rmv_page_order(struct page *page)
477 __ClearPageBuddy(page);
478 set_page_private(page, 0);
482 * This function checks whether a page is free && is the buddy
483 * we can do coalesce a page and its buddy if
484 * (a) the buddy is not in a hole &&
485 * (b) the buddy is in the buddy system &&
486 * (c) a page and its buddy have the same order &&
487 * (d) a page and its buddy are in the same zone.
489 * For recording whether a page is in the buddy system, we set ->_mapcount
490 * PAGE_BUDDY_MAPCOUNT_VALUE.
491 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
492 * serialized by zone->lock.
494 * For recording page's order, we use page_private(page).
496 static inline int page_is_buddy(struct page *page, struct page *buddy,
499 if (!pfn_valid_within(page_to_pfn(buddy)))
502 if (page_is_guard(buddy) && page_order(buddy) == order) {
503 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
505 if (page_zone_id(page) != page_zone_id(buddy))
511 if (PageBuddy(buddy) && page_order(buddy) == order) {
512 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
515 * zone check is done late to avoid uselessly
516 * calculating zone/node ids for pages that could
519 if (page_zone_id(page) != page_zone_id(buddy))
528 * Freeing function for a buddy system allocator.
530 * The concept of a buddy system is to maintain direct-mapped table
531 * (containing bit values) for memory blocks of various "orders".
532 * The bottom level table contains the map for the smallest allocatable
533 * units of memory (here, pages), and each level above it describes
534 * pairs of units from the levels below, hence, "buddies".
535 * At a high level, all that happens here is marking the table entry
536 * at the bottom level available, and propagating the changes upward
537 * as necessary, plus some accounting needed to play nicely with other
538 * parts of the VM system.
539 * At each level, we keep a list of pages, which are heads of continuous
540 * free pages of length of (1 << order) and marked with _mapcount
541 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
543 * So when we are allocating or freeing one, we can derive the state of the
544 * other. That is, if we allocate a small block, and both were
545 * free, the remainder of the region must be split into blocks.
546 * If a block is freed, and its buddy is also free, then this
547 * triggers coalescing into a block of larger size.
552 static inline void __free_one_page(struct page *page,
554 struct zone *zone, unsigned int order,
557 unsigned long page_idx;
558 unsigned long combined_idx;
559 unsigned long uninitialized_var(buddy_idx);
561 int max_order = MAX_ORDER;
563 VM_BUG_ON(!zone_is_initialized(zone));
565 if (unlikely(PageCompound(page)))
566 if (unlikely(destroy_compound_page(page, order)))
569 VM_BUG_ON(migratetype == -1);
570 if (is_migrate_isolate(migratetype)) {
572 * We restrict max order of merging to prevent merge
573 * between freepages on isolate pageblock and normal
574 * pageblock. Without this, pageblock isolation
575 * could cause incorrect freepage accounting.
577 max_order = min(MAX_ORDER, pageblock_order + 1);
579 __mod_zone_freepage_state(zone, 1 << order, migratetype);
582 page_idx = pfn & ((1 << max_order) - 1);
584 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
585 VM_BUG_ON_PAGE(bad_range(zone, page), page);
587 while (order < max_order - 1) {
588 buddy_idx = __find_buddy_index(page_idx, order);
589 buddy = page + (buddy_idx - page_idx);
590 if (!page_is_buddy(page, buddy, order))
593 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
594 * merge with it and move up one order.
596 if (page_is_guard(buddy)) {
597 clear_page_guard_flag(buddy);
598 set_page_private(buddy, 0);
599 if (!is_migrate_isolate(migratetype)) {
600 __mod_zone_freepage_state(zone, 1 << order,
604 list_del(&buddy->lru);
605 zone->free_area[order].nr_free--;
606 rmv_page_order(buddy);
608 combined_idx = buddy_idx & page_idx;
609 page = page + (combined_idx - page_idx);
610 page_idx = combined_idx;
613 set_page_order(page, order);
616 * If this is not the largest possible page, check if the buddy
617 * of the next-highest order is free. If it is, it's possible
618 * that pages are being freed that will coalesce soon. In case,
619 * that is happening, add the free page to the tail of the list
620 * so it's less likely to be used soon and more likely to be merged
621 * as a higher order page
623 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
624 struct page *higher_page, *higher_buddy;
625 combined_idx = buddy_idx & page_idx;
626 higher_page = page + (combined_idx - page_idx);
627 buddy_idx = __find_buddy_index(combined_idx, order + 1);
628 higher_buddy = higher_page + (buddy_idx - combined_idx);
629 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
630 list_add_tail(&page->lru,
631 &zone->free_area[order].free_list[migratetype]);
636 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
638 zone->free_area[order].nr_free++;
641 static inline int free_pages_check(struct page *page)
643 const char *bad_reason = NULL;
644 unsigned long bad_flags = 0;
646 if (unlikely(page_mapcount(page)))
647 bad_reason = "nonzero mapcount";
648 if (unlikely(page->mapping != NULL))
649 bad_reason = "non-NULL mapping";
650 if (unlikely(atomic_read(&page->_count) != 0))
651 bad_reason = "nonzero _count";
652 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
653 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
654 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
656 if (unlikely(mem_cgroup_bad_page_check(page)))
657 bad_reason = "cgroup check failed";
658 if (unlikely(bad_reason)) {
659 bad_page(page, bad_reason, bad_flags);
662 page_cpupid_reset_last(page);
663 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
664 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
669 * Frees a number of pages from the PCP lists
670 * Assumes all pages on list are in same zone, and of same order.
671 * count is the number of pages to free.
673 * If the zone was previously in an "all pages pinned" state then look to
674 * see if this freeing clears that state.
676 * And clear the zone's pages_scanned counter, to hold off the "all pages are
677 * pinned" detection logic.
679 static void free_pcppages_bulk(struct zone *zone, int count,
680 struct per_cpu_pages *pcp)
685 unsigned long nr_scanned;
687 spin_lock(&zone->lock);
688 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
690 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
694 struct list_head *list;
697 * Remove pages from lists in a round-robin fashion. A
698 * batch_free count is maintained that is incremented when an
699 * empty list is encountered. This is so more pages are freed
700 * off fuller lists instead of spinning excessively around empty
705 if (++migratetype == MIGRATE_PCPTYPES)
707 list = &pcp->lists[migratetype];
708 } while (list_empty(list));
710 /* This is the only non-empty list. Free them all. */
711 if (batch_free == MIGRATE_PCPTYPES)
712 batch_free = to_free;
715 int mt; /* migratetype of the to-be-freed page */
717 page = list_entry(list->prev, struct page, lru);
718 /* must delete as __free_one_page list manipulates */
719 list_del(&page->lru);
720 mt = get_freepage_migratetype(page);
721 if (unlikely(has_isolate_pageblock(zone)))
722 mt = get_pageblock_migratetype(page);
724 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
725 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
726 trace_mm_page_pcpu_drain(page, 0, mt);
727 } while (--to_free && --batch_free && !list_empty(list));
729 spin_unlock(&zone->lock);
732 static void free_one_page(struct zone *zone,
733 struct page *page, unsigned long pfn,
737 unsigned long nr_scanned;
738 spin_lock(&zone->lock);
739 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
741 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
743 if (unlikely(has_isolate_pageblock(zone) ||
744 is_migrate_isolate(migratetype))) {
745 migratetype = get_pfnblock_migratetype(page, pfn);
747 __free_one_page(page, pfn, zone, order, migratetype);
748 spin_unlock(&zone->lock);
751 static bool free_pages_prepare(struct page *page, unsigned int order)
756 trace_mm_page_free(page, order);
757 kmemcheck_free_shadow(page, order);
760 page->mapping = NULL;
761 for (i = 0; i < (1 << order); i++)
762 bad += free_pages_check(page + i);
766 if (!PageHighMem(page)) {
767 debug_check_no_locks_freed(page_address(page),
769 debug_check_no_obj_freed(page_address(page),
772 arch_free_page(page, order);
773 kernel_map_pages(page, 1 << order, 0);
778 static void __free_pages_ok(struct page *page, unsigned int order)
782 unsigned long pfn = page_to_pfn(page);
784 if (!free_pages_prepare(page, order))
787 migratetype = get_pfnblock_migratetype(page, pfn);
788 local_irq_save(flags);
789 __count_vm_events(PGFREE, 1 << order);
790 set_freepage_migratetype(page, migratetype);
791 free_one_page(page_zone(page), page, pfn, order, migratetype);
792 local_irq_restore(flags);
795 void __init __free_pages_bootmem(struct page *page, unsigned int order)
797 unsigned int nr_pages = 1 << order;
798 struct page *p = page;
802 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
804 __ClearPageReserved(p);
805 set_page_count(p, 0);
807 __ClearPageReserved(p);
808 set_page_count(p, 0);
810 page_zone(page)->managed_pages += nr_pages;
811 set_page_refcounted(page);
812 __free_pages(page, order);
816 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
817 void __init init_cma_reserved_pageblock(struct page *page)
819 unsigned i = pageblock_nr_pages;
820 struct page *p = page;
823 __ClearPageReserved(p);
824 set_page_count(p, 0);
827 set_pageblock_migratetype(page, MIGRATE_CMA);
829 if (pageblock_order >= MAX_ORDER) {
830 i = pageblock_nr_pages;
833 set_page_refcounted(p);
834 __free_pages(p, MAX_ORDER - 1);
835 p += MAX_ORDER_NR_PAGES;
836 } while (i -= MAX_ORDER_NR_PAGES);
838 set_page_refcounted(page);
839 __free_pages(page, pageblock_order);
842 adjust_managed_page_count(page, pageblock_nr_pages);
847 * The order of subdivision here is critical for the IO subsystem.
848 * Please do not alter this order without good reasons and regression
849 * testing. Specifically, as large blocks of memory are subdivided,
850 * the order in which smaller blocks are delivered depends on the order
851 * they're subdivided in this function. This is the primary factor
852 * influencing the order in which pages are delivered to the IO
853 * subsystem according to empirical testing, and this is also justified
854 * by considering the behavior of a buddy system containing a single
855 * large block of memory acted on by a series of small allocations.
856 * This behavior is a critical factor in sglist merging's success.
860 static inline void expand(struct zone *zone, struct page *page,
861 int low, int high, struct free_area *area,
864 unsigned long size = 1 << high;
870 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
872 #ifdef CONFIG_DEBUG_PAGEALLOC
873 if (high < debug_guardpage_minorder()) {
875 * Mark as guard pages (or page), that will allow to
876 * merge back to allocator when buddy will be freed.
877 * Corresponding page table entries will not be touched,
878 * pages will stay not present in virtual address space
880 INIT_LIST_HEAD(&page[size].lru);
881 set_page_guard_flag(&page[size]);
882 set_page_private(&page[size], high);
883 /* Guard pages are not available for any usage */
884 __mod_zone_freepage_state(zone, -(1 << high),
889 list_add(&page[size].lru, &area->free_list[migratetype]);
891 set_page_order(&page[size], high);
896 * This page is about to be returned from the page allocator
898 static inline int check_new_page(struct page *page)
900 const char *bad_reason = NULL;
901 unsigned long bad_flags = 0;
903 if (unlikely(page_mapcount(page)))
904 bad_reason = "nonzero mapcount";
905 if (unlikely(page->mapping != NULL))
906 bad_reason = "non-NULL mapping";
907 if (unlikely(atomic_read(&page->_count) != 0))
908 bad_reason = "nonzero _count";
909 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
910 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
911 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
913 if (unlikely(mem_cgroup_bad_page_check(page)))
914 bad_reason = "cgroup check failed";
915 if (unlikely(bad_reason)) {
916 bad_page(page, bad_reason, bad_flags);
922 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
926 for (i = 0; i < (1 << order); i++) {
927 struct page *p = page + i;
928 if (unlikely(check_new_page(p)))
932 set_page_private(page, 0);
933 set_page_refcounted(page);
935 arch_alloc_page(page, order);
936 kernel_map_pages(page, 1 << order, 1);
938 if (gfp_flags & __GFP_ZERO)
939 prep_zero_page(page, order, gfp_flags);
941 if (order && (gfp_flags & __GFP_COMP))
942 prep_compound_page(page, order);
948 * Go through the free lists for the given migratetype and remove
949 * the smallest available page from the freelists
952 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
955 unsigned int current_order;
956 struct free_area *area;
959 /* Find a page of the appropriate size in the preferred list */
960 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
961 area = &(zone->free_area[current_order]);
962 if (list_empty(&area->free_list[migratetype]))
965 page = list_entry(area->free_list[migratetype].next,
967 list_del(&page->lru);
968 rmv_page_order(page);
970 expand(zone, page, order, current_order, area, migratetype);
971 set_freepage_migratetype(page, migratetype);
980 * This array describes the order lists are fallen back to when
981 * the free lists for the desirable migrate type are depleted
983 static int fallbacks[MIGRATE_TYPES][4] = {
984 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
985 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
987 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
988 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
990 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
992 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
993 #ifdef CONFIG_MEMORY_ISOLATION
994 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
999 * Move the free pages in a range to the free lists of the requested type.
1000 * Note that start_page and end_pages are not aligned on a pageblock
1001 * boundary. If alignment is required, use move_freepages_block()
1003 int move_freepages(struct zone *zone,
1004 struct page *start_page, struct page *end_page,
1008 unsigned long order;
1009 int pages_moved = 0;
1011 #ifndef CONFIG_HOLES_IN_ZONE
1013 * page_zone is not safe to call in this context when
1014 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1015 * anyway as we check zone boundaries in move_freepages_block().
1016 * Remove at a later date when no bug reports exist related to
1017 * grouping pages by mobility
1019 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1022 for (page = start_page; page <= end_page;) {
1023 /* Make sure we are not inadvertently changing nodes */
1024 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1026 if (!pfn_valid_within(page_to_pfn(page))) {
1031 if (!PageBuddy(page)) {
1036 order = page_order(page);
1037 list_move(&page->lru,
1038 &zone->free_area[order].free_list[migratetype]);
1039 set_freepage_migratetype(page, migratetype);
1041 pages_moved += 1 << order;
1047 int move_freepages_block(struct zone *zone, struct page *page,
1050 unsigned long start_pfn, end_pfn;
1051 struct page *start_page, *end_page;
1053 start_pfn = page_to_pfn(page);
1054 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1055 start_page = pfn_to_page(start_pfn);
1056 end_page = start_page + pageblock_nr_pages - 1;
1057 end_pfn = start_pfn + pageblock_nr_pages - 1;
1059 /* Do not cross zone boundaries */
1060 if (!zone_spans_pfn(zone, start_pfn))
1062 if (!zone_spans_pfn(zone, end_pfn))
1065 return move_freepages(zone, start_page, end_page, migratetype);
1068 static void change_pageblock_range(struct page *pageblock_page,
1069 int start_order, int migratetype)
1071 int nr_pageblocks = 1 << (start_order - pageblock_order);
1073 while (nr_pageblocks--) {
1074 set_pageblock_migratetype(pageblock_page, migratetype);
1075 pageblock_page += pageblock_nr_pages;
1080 * If breaking a large block of pages, move all free pages to the preferred
1081 * allocation list. If falling back for a reclaimable kernel allocation, be
1082 * more aggressive about taking ownership of free pages.
1084 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1085 * nor move CMA pages to different free lists. We don't want unmovable pages
1086 * to be allocated from MIGRATE_CMA areas.
1088 * Returns the new migratetype of the pageblock (or the same old migratetype
1089 * if it was unchanged).
1091 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1092 int start_type, int fallback_type)
1094 int current_order = page_order(page);
1097 * When borrowing from MIGRATE_CMA, we need to release the excess
1098 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1099 * is set to CMA so it is returned to the correct freelist in case
1100 * the page ends up being not actually allocated from the pcp lists.
1102 if (is_migrate_cma(fallback_type))
1103 return fallback_type;
1105 /* Take ownership for orders >= pageblock_order */
1106 if (current_order >= pageblock_order) {
1107 change_pageblock_range(page, current_order, start_type);
1111 if (current_order >= pageblock_order / 2 ||
1112 start_type == MIGRATE_RECLAIMABLE ||
1113 page_group_by_mobility_disabled) {
1116 pages = move_freepages_block(zone, page, start_type);
1118 /* Claim the whole block if over half of it is free */
1119 if (pages >= (1 << (pageblock_order-1)) ||
1120 page_group_by_mobility_disabled) {
1122 set_pageblock_migratetype(page, start_type);
1128 return fallback_type;
1131 /* Remove an element from the buddy allocator from the fallback list */
1132 static inline struct page *
1133 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1135 struct free_area *area;
1136 unsigned int current_order;
1138 int migratetype, new_type, i;
1140 /* Find the largest possible block of pages in the other list */
1141 for (current_order = MAX_ORDER-1;
1142 current_order >= order && current_order <= MAX_ORDER-1;
1145 migratetype = fallbacks[start_migratetype][i];
1147 /* MIGRATE_RESERVE handled later if necessary */
1148 if (migratetype == MIGRATE_RESERVE)
1151 area = &(zone->free_area[current_order]);
1152 if (list_empty(&area->free_list[migratetype]))
1155 page = list_entry(area->free_list[migratetype].next,
1159 new_type = try_to_steal_freepages(zone, page,
1163 /* Remove the page from the freelists */
1164 list_del(&page->lru);
1165 rmv_page_order(page);
1167 expand(zone, page, order, current_order, area,
1169 /* The freepage_migratetype may differ from pageblock's
1170 * migratetype depending on the decisions in
1171 * try_to_steal_freepages. This is OK as long as it does
1172 * not differ for MIGRATE_CMA type.
1174 set_freepage_migratetype(page, new_type);
1176 trace_mm_page_alloc_extfrag(page, order, current_order,
1177 start_migratetype, migratetype, new_type);
1187 * Do the hard work of removing an element from the buddy allocator.
1188 * Call me with the zone->lock already held.
1190 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1196 page = __rmqueue_smallest(zone, order, migratetype);
1198 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1199 page = __rmqueue_fallback(zone, order, migratetype);
1202 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1203 * is used because __rmqueue_smallest is an inline function
1204 * and we want just one call site
1207 migratetype = MIGRATE_RESERVE;
1212 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1217 * Obtain a specified number of elements from the buddy allocator, all under
1218 * a single hold of the lock, for efficiency. Add them to the supplied list.
1219 * Returns the number of new pages which were placed at *list.
1221 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1222 unsigned long count, struct list_head *list,
1223 int migratetype, bool cold)
1227 spin_lock(&zone->lock);
1228 for (i = 0; i < count; ++i) {
1229 struct page *page = __rmqueue(zone, order, migratetype);
1230 if (unlikely(page == NULL))
1234 * Split buddy pages returned by expand() are received here
1235 * in physical page order. The page is added to the callers and
1236 * list and the list head then moves forward. From the callers
1237 * perspective, the linked list is ordered by page number in
1238 * some conditions. This is useful for IO devices that can
1239 * merge IO requests if the physical pages are ordered
1243 list_add(&page->lru, list);
1245 list_add_tail(&page->lru, list);
1247 if (is_migrate_cma(get_freepage_migratetype(page)))
1248 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1251 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1252 spin_unlock(&zone->lock);
1258 * Called from the vmstat counter updater to drain pagesets of this
1259 * currently executing processor on remote nodes after they have
1262 * Note that this function must be called with the thread pinned to
1263 * a single processor.
1265 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1267 unsigned long flags;
1268 int to_drain, batch;
1270 local_irq_save(flags);
1271 batch = ACCESS_ONCE(pcp->batch);
1272 to_drain = min(pcp->count, batch);
1274 free_pcppages_bulk(zone, to_drain, pcp);
1275 pcp->count -= to_drain;
1277 local_irq_restore(flags);
1282 * Drain pages of the indicated processor.
1284 * The processor must either be the current processor and the
1285 * thread pinned to the current processor or a processor that
1288 static void drain_pages(unsigned int cpu)
1290 unsigned long flags;
1293 for_each_populated_zone(zone) {
1294 struct per_cpu_pageset *pset;
1295 struct per_cpu_pages *pcp;
1297 local_irq_save(flags);
1298 pset = per_cpu_ptr(zone->pageset, cpu);
1302 free_pcppages_bulk(zone, pcp->count, pcp);
1305 local_irq_restore(flags);
1310 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1312 void drain_local_pages(void *arg)
1314 drain_pages(smp_processor_id());
1318 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1320 * Note that this code is protected against sending an IPI to an offline
1321 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1322 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1323 * nothing keeps CPUs from showing up after we populated the cpumask and
1324 * before the call to on_each_cpu_mask().
1326 void drain_all_pages(void)
1329 struct per_cpu_pageset *pcp;
1333 * Allocate in the BSS so we wont require allocation in
1334 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1336 static cpumask_t cpus_with_pcps;
1339 * We don't care about racing with CPU hotplug event
1340 * as offline notification will cause the notified
1341 * cpu to drain that CPU pcps and on_each_cpu_mask
1342 * disables preemption as part of its processing
1344 for_each_online_cpu(cpu) {
1345 bool has_pcps = false;
1346 for_each_populated_zone(zone) {
1347 pcp = per_cpu_ptr(zone->pageset, cpu);
1348 if (pcp->pcp.count) {
1354 cpumask_set_cpu(cpu, &cpus_with_pcps);
1356 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1358 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1361 #ifdef CONFIG_HIBERNATION
1363 void mark_free_pages(struct zone *zone)
1365 unsigned long pfn, max_zone_pfn;
1366 unsigned long flags;
1367 unsigned int order, t;
1368 struct list_head *curr;
1370 if (zone_is_empty(zone))
1373 spin_lock_irqsave(&zone->lock, flags);
1375 max_zone_pfn = zone_end_pfn(zone);
1376 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1377 if (pfn_valid(pfn)) {
1378 struct page *page = pfn_to_page(pfn);
1380 if (!swsusp_page_is_forbidden(page))
1381 swsusp_unset_page_free(page);
1384 for_each_migratetype_order(order, t) {
1385 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1388 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1389 for (i = 0; i < (1UL << order); i++)
1390 swsusp_set_page_free(pfn_to_page(pfn + i));
1393 spin_unlock_irqrestore(&zone->lock, flags);
1395 #endif /* CONFIG_PM */
1398 * Free a 0-order page
1399 * cold == true ? free a cold page : free a hot page
1401 void free_hot_cold_page(struct page *page, bool cold)
1403 struct zone *zone = page_zone(page);
1404 struct per_cpu_pages *pcp;
1405 unsigned long flags;
1406 unsigned long pfn = page_to_pfn(page);
1409 if (!free_pages_prepare(page, 0))
1412 migratetype = get_pfnblock_migratetype(page, pfn);
1413 set_freepage_migratetype(page, migratetype);
1414 local_irq_save(flags);
1415 __count_vm_event(PGFREE);
1418 * We only track unmovable, reclaimable and movable on pcp lists.
1419 * Free ISOLATE pages back to the allocator because they are being
1420 * offlined but treat RESERVE as movable pages so we can get those
1421 * areas back if necessary. Otherwise, we may have to free
1422 * excessively into the page allocator
1424 if (migratetype >= MIGRATE_PCPTYPES) {
1425 if (unlikely(is_migrate_isolate(migratetype))) {
1426 free_one_page(zone, page, pfn, 0, migratetype);
1429 migratetype = MIGRATE_MOVABLE;
1432 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1434 list_add(&page->lru, &pcp->lists[migratetype]);
1436 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1438 if (pcp->count >= pcp->high) {
1439 unsigned long batch = ACCESS_ONCE(pcp->batch);
1440 free_pcppages_bulk(zone, batch, pcp);
1441 pcp->count -= batch;
1445 local_irq_restore(flags);
1449 * Free a list of 0-order pages
1451 void free_hot_cold_page_list(struct list_head *list, bool cold)
1453 struct page *page, *next;
1455 list_for_each_entry_safe(page, next, list, lru) {
1456 trace_mm_page_free_batched(page, cold);
1457 free_hot_cold_page(page, cold);
1462 * split_page takes a non-compound higher-order page, and splits it into
1463 * n (1<<order) sub-pages: page[0..n]
1464 * Each sub-page must be freed individually.
1466 * Note: this is probably too low level an operation for use in drivers.
1467 * Please consult with lkml before using this in your driver.
1469 void split_page(struct page *page, unsigned int order)
1473 VM_BUG_ON_PAGE(PageCompound(page), page);
1474 VM_BUG_ON_PAGE(!page_count(page), page);
1476 #ifdef CONFIG_KMEMCHECK
1478 * Split shadow pages too, because free(page[0]) would
1479 * otherwise free the whole shadow.
1481 if (kmemcheck_page_is_tracked(page))
1482 split_page(virt_to_page(page[0].shadow), order);
1485 for (i = 1; i < (1 << order); i++)
1486 set_page_refcounted(page + i);
1488 EXPORT_SYMBOL_GPL(split_page);
1490 int __isolate_free_page(struct page *page, unsigned int order)
1492 unsigned long watermark;
1496 BUG_ON(!PageBuddy(page));
1498 zone = page_zone(page);
1499 mt = get_pageblock_migratetype(page);
1501 if (!is_migrate_isolate(mt)) {
1502 /* Obey watermarks as if the page was being allocated */
1503 watermark = low_wmark_pages(zone) + (1 << order);
1504 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1507 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1510 /* Remove page from free list */
1511 list_del(&page->lru);
1512 zone->free_area[order].nr_free--;
1513 rmv_page_order(page);
1515 /* Set the pageblock if the isolated page is at least a pageblock */
1516 if (order >= pageblock_order - 1) {
1517 struct page *endpage = page + (1 << order) - 1;
1518 for (; page < endpage; page += pageblock_nr_pages) {
1519 int mt = get_pageblock_migratetype(page);
1520 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1521 set_pageblock_migratetype(page,
1526 return 1UL << order;
1530 * Similar to split_page except the page is already free. As this is only
1531 * being used for migration, the migratetype of the block also changes.
1532 * As this is called with interrupts disabled, the caller is responsible
1533 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1536 * Note: this is probably too low level an operation for use in drivers.
1537 * Please consult with lkml before using this in your driver.
1539 int split_free_page(struct page *page)
1544 order = page_order(page);
1546 nr_pages = __isolate_free_page(page, order);
1550 /* Split into individual pages */
1551 set_page_refcounted(page);
1552 split_page(page, order);
1557 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1558 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1562 struct page *buffered_rmqueue(struct zone *preferred_zone,
1563 struct zone *zone, unsigned int order,
1564 gfp_t gfp_flags, int migratetype)
1566 unsigned long flags;
1568 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1571 if (likely(order == 0)) {
1572 struct per_cpu_pages *pcp;
1573 struct list_head *list;
1575 local_irq_save(flags);
1576 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1577 list = &pcp->lists[migratetype];
1578 if (list_empty(list)) {
1579 pcp->count += rmqueue_bulk(zone, 0,
1582 if (unlikely(list_empty(list)))
1587 page = list_entry(list->prev, struct page, lru);
1589 page = list_entry(list->next, struct page, lru);
1591 list_del(&page->lru);
1594 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1596 * __GFP_NOFAIL is not to be used in new code.
1598 * All __GFP_NOFAIL callers should be fixed so that they
1599 * properly detect and handle allocation failures.
1601 * We most definitely don't want callers attempting to
1602 * allocate greater than order-1 page units with
1605 WARN_ON_ONCE(order > 1);
1607 spin_lock_irqsave(&zone->lock, flags);
1608 page = __rmqueue(zone, order, migratetype);
1609 spin_unlock(&zone->lock);
1612 __mod_zone_freepage_state(zone, -(1 << order),
1613 get_freepage_migratetype(page));
1616 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1617 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1618 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1619 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1621 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1622 zone_statistics(preferred_zone, zone, gfp_flags);
1623 local_irq_restore(flags);
1625 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1626 if (prep_new_page(page, order, gfp_flags))
1631 local_irq_restore(flags);
1635 #ifdef CONFIG_FAIL_PAGE_ALLOC
1638 struct fault_attr attr;
1640 u32 ignore_gfp_highmem;
1641 u32 ignore_gfp_wait;
1643 } fail_page_alloc = {
1644 .attr = FAULT_ATTR_INITIALIZER,
1645 .ignore_gfp_wait = 1,
1646 .ignore_gfp_highmem = 1,
1650 static int __init setup_fail_page_alloc(char *str)
1652 return setup_fault_attr(&fail_page_alloc.attr, str);
1654 __setup("fail_page_alloc=", setup_fail_page_alloc);
1656 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1658 if (order < fail_page_alloc.min_order)
1660 if (gfp_mask & __GFP_NOFAIL)
1662 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1664 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1667 return should_fail(&fail_page_alloc.attr, 1 << order);
1670 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1672 static int __init fail_page_alloc_debugfs(void)
1674 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1677 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1678 &fail_page_alloc.attr);
1680 return PTR_ERR(dir);
1682 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1683 &fail_page_alloc.ignore_gfp_wait))
1685 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1686 &fail_page_alloc.ignore_gfp_highmem))
1688 if (!debugfs_create_u32("min-order", mode, dir,
1689 &fail_page_alloc.min_order))
1694 debugfs_remove_recursive(dir);
1699 late_initcall(fail_page_alloc_debugfs);
1701 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1703 #else /* CONFIG_FAIL_PAGE_ALLOC */
1705 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1710 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1713 * Return true if free pages are above 'mark'. This takes into account the order
1714 * of the allocation.
1716 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1717 unsigned long mark, int classzone_idx, int alloc_flags,
1720 /* free_pages my go negative - that's OK */
1725 free_pages -= (1 << order) - 1;
1726 if (alloc_flags & ALLOC_HIGH)
1728 if (alloc_flags & ALLOC_HARDER)
1731 /* If allocation can't use CMA areas don't use free CMA pages */
1732 if (!(alloc_flags & ALLOC_CMA))
1733 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1736 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1738 for (o = 0; o < order; o++) {
1739 /* At the next order, this order's pages become unavailable */
1740 free_pages -= z->free_area[o].nr_free << o;
1742 /* Require fewer higher order pages to be free */
1745 if (free_pages <= min)
1751 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1752 int classzone_idx, int alloc_flags)
1754 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1755 zone_page_state(z, NR_FREE_PAGES));
1758 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1759 unsigned long mark, int classzone_idx, int alloc_flags)
1761 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1763 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1764 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1766 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1772 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1773 * skip over zones that are not allowed by the cpuset, or that have
1774 * been recently (in last second) found to be nearly full. See further
1775 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1776 * that have to skip over a lot of full or unallowed zones.
1778 * If the zonelist cache is present in the passed zonelist, then
1779 * returns a pointer to the allowed node mask (either the current
1780 * tasks mems_allowed, or node_states[N_MEMORY].)
1782 * If the zonelist cache is not available for this zonelist, does
1783 * nothing and returns NULL.
1785 * If the fullzones BITMAP in the zonelist cache is stale (more than
1786 * a second since last zap'd) then we zap it out (clear its bits.)
1788 * We hold off even calling zlc_setup, until after we've checked the
1789 * first zone in the zonelist, on the theory that most allocations will
1790 * be satisfied from that first zone, so best to examine that zone as
1791 * quickly as we can.
1793 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1795 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1796 nodemask_t *allowednodes; /* zonelist_cache approximation */
1798 zlc = zonelist->zlcache_ptr;
1802 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1803 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1804 zlc->last_full_zap = jiffies;
1807 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1808 &cpuset_current_mems_allowed :
1809 &node_states[N_MEMORY];
1810 return allowednodes;
1814 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1815 * if it is worth looking at further for free memory:
1816 * 1) Check that the zone isn't thought to be full (doesn't have its
1817 * bit set in the zonelist_cache fullzones BITMAP).
1818 * 2) Check that the zones node (obtained from the zonelist_cache
1819 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1820 * Return true (non-zero) if zone is worth looking at further, or
1821 * else return false (zero) if it is not.
1823 * This check -ignores- the distinction between various watermarks,
1824 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1825 * found to be full for any variation of these watermarks, it will
1826 * be considered full for up to one second by all requests, unless
1827 * we are so low on memory on all allowed nodes that we are forced
1828 * into the second scan of the zonelist.
1830 * In the second scan we ignore this zonelist cache and exactly
1831 * apply the watermarks to all zones, even it is slower to do so.
1832 * We are low on memory in the second scan, and should leave no stone
1833 * unturned looking for a free page.
1835 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1836 nodemask_t *allowednodes)
1838 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1839 int i; /* index of *z in zonelist zones */
1840 int n; /* node that zone *z is on */
1842 zlc = zonelist->zlcache_ptr;
1846 i = z - zonelist->_zonerefs;
1849 /* This zone is worth trying if it is allowed but not full */
1850 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1854 * Given 'z' scanning a zonelist, set the corresponding bit in
1855 * zlc->fullzones, so that subsequent attempts to allocate a page
1856 * from that zone don't waste time re-examining it.
1858 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1860 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1861 int i; /* index of *z in zonelist zones */
1863 zlc = zonelist->zlcache_ptr;
1867 i = z - zonelist->_zonerefs;
1869 set_bit(i, zlc->fullzones);
1873 * clear all zones full, called after direct reclaim makes progress so that
1874 * a zone that was recently full is not skipped over for up to a second
1876 static void zlc_clear_zones_full(struct zonelist *zonelist)
1878 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1880 zlc = zonelist->zlcache_ptr;
1884 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1887 static bool zone_local(struct zone *local_zone, struct zone *zone)
1889 return local_zone->node == zone->node;
1892 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1894 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1898 #else /* CONFIG_NUMA */
1900 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1905 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1906 nodemask_t *allowednodes)
1911 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1915 static void zlc_clear_zones_full(struct zonelist *zonelist)
1919 static bool zone_local(struct zone *local_zone, struct zone *zone)
1924 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1929 #endif /* CONFIG_NUMA */
1931 static void reset_alloc_batches(struct zone *preferred_zone)
1933 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
1936 mod_zone_page_state(zone, NR_ALLOC_BATCH,
1937 high_wmark_pages(zone) - low_wmark_pages(zone) -
1938 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
1939 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1940 } while (zone++ != preferred_zone);
1944 * get_page_from_freelist goes through the zonelist trying to allocate
1947 static struct page *
1948 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1949 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1950 struct zone *preferred_zone, int classzone_idx, int migratetype)
1953 struct page *page = NULL;
1955 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1956 int zlc_active = 0; /* set if using zonelist_cache */
1957 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1958 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1959 (gfp_mask & __GFP_WRITE);
1960 int nr_fair_skipped = 0;
1961 bool zonelist_rescan;
1964 zonelist_rescan = false;
1967 * Scan zonelist, looking for a zone with enough free.
1968 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1970 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1971 high_zoneidx, nodemask) {
1974 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1975 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1977 if (cpusets_enabled() &&
1978 (alloc_flags & ALLOC_CPUSET) &&
1979 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1982 * Distribute pages in proportion to the individual
1983 * zone size to ensure fair page aging. The zone a
1984 * page was allocated in should have no effect on the
1985 * time the page has in memory before being reclaimed.
1987 if (alloc_flags & ALLOC_FAIR) {
1988 if (!zone_local(preferred_zone, zone))
1990 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
1996 * When allocating a page cache page for writing, we
1997 * want to get it from a zone that is within its dirty
1998 * limit, such that no single zone holds more than its
1999 * proportional share of globally allowed dirty pages.
2000 * The dirty limits take into account the zone's
2001 * lowmem reserves and high watermark so that kswapd
2002 * should be able to balance it without having to
2003 * write pages from its LRU list.
2005 * This may look like it could increase pressure on
2006 * lower zones by failing allocations in higher zones
2007 * before they are full. But the pages that do spill
2008 * over are limited as the lower zones are protected
2009 * by this very same mechanism. It should not become
2010 * a practical burden to them.
2012 * XXX: For now, allow allocations to potentially
2013 * exceed the per-zone dirty limit in the slowpath
2014 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2015 * which is important when on a NUMA setup the allowed
2016 * zones are together not big enough to reach the
2017 * global limit. The proper fix for these situations
2018 * will require awareness of zones in the
2019 * dirty-throttling and the flusher threads.
2021 if (consider_zone_dirty && !zone_dirty_ok(zone))
2024 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2025 if (!zone_watermark_ok(zone, order, mark,
2026 classzone_idx, alloc_flags)) {
2029 /* Checked here to keep the fast path fast */
2030 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2031 if (alloc_flags & ALLOC_NO_WATERMARKS)
2034 if (IS_ENABLED(CONFIG_NUMA) &&
2035 !did_zlc_setup && nr_online_nodes > 1) {
2037 * we do zlc_setup if there are multiple nodes
2038 * and before considering the first zone allowed
2041 allowednodes = zlc_setup(zonelist, alloc_flags);
2046 if (zone_reclaim_mode == 0 ||
2047 !zone_allows_reclaim(preferred_zone, zone))
2048 goto this_zone_full;
2051 * As we may have just activated ZLC, check if the first
2052 * eligible zone has failed zone_reclaim recently.
2054 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2055 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2058 ret = zone_reclaim(zone, gfp_mask, order);
2060 case ZONE_RECLAIM_NOSCAN:
2063 case ZONE_RECLAIM_FULL:
2064 /* scanned but unreclaimable */
2067 /* did we reclaim enough */
2068 if (zone_watermark_ok(zone, order, mark,
2069 classzone_idx, alloc_flags))
2073 * Failed to reclaim enough to meet watermark.
2074 * Only mark the zone full if checking the min
2075 * watermark or if we failed to reclaim just
2076 * 1<<order pages or else the page allocator
2077 * fastpath will prematurely mark zones full
2078 * when the watermark is between the low and
2081 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2082 ret == ZONE_RECLAIM_SOME)
2083 goto this_zone_full;
2090 page = buffered_rmqueue(preferred_zone, zone, order,
2091 gfp_mask, migratetype);
2095 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2096 zlc_mark_zone_full(zonelist, z);
2101 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2102 * necessary to allocate the page. The expectation is
2103 * that the caller is taking steps that will free more
2104 * memory. The caller should avoid the page being used
2105 * for !PFMEMALLOC purposes.
2107 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2112 * The first pass makes sure allocations are spread fairly within the
2113 * local node. However, the local node might have free pages left
2114 * after the fairness batches are exhausted, and remote zones haven't
2115 * even been considered yet. Try once more without fairness, and
2116 * include remote zones now, before entering the slowpath and waking
2117 * kswapd: prefer spilling to a remote zone over swapping locally.
2119 if (alloc_flags & ALLOC_FAIR) {
2120 alloc_flags &= ~ALLOC_FAIR;
2121 if (nr_fair_skipped) {
2122 zonelist_rescan = true;
2123 reset_alloc_batches(preferred_zone);
2125 if (nr_online_nodes > 1)
2126 zonelist_rescan = true;
2129 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2130 /* Disable zlc cache for second zonelist scan */
2132 zonelist_rescan = true;
2135 if (zonelist_rescan)
2142 * Large machines with many possible nodes should not always dump per-node
2143 * meminfo in irq context.
2145 static inline bool should_suppress_show_mem(void)
2150 ret = in_interrupt();
2155 static DEFINE_RATELIMIT_STATE(nopage_rs,
2156 DEFAULT_RATELIMIT_INTERVAL,
2157 DEFAULT_RATELIMIT_BURST);
2159 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2161 unsigned int filter = SHOW_MEM_FILTER_NODES;
2163 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2164 debug_guardpage_minorder() > 0)
2168 * This documents exceptions given to allocations in certain
2169 * contexts that are allowed to allocate outside current's set
2172 if (!(gfp_mask & __GFP_NOMEMALLOC))
2173 if (test_thread_flag(TIF_MEMDIE) ||
2174 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2175 filter &= ~SHOW_MEM_FILTER_NODES;
2176 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2177 filter &= ~SHOW_MEM_FILTER_NODES;
2180 struct va_format vaf;
2183 va_start(args, fmt);
2188 pr_warn("%pV", &vaf);
2193 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2194 current->comm, order, gfp_mask);
2197 if (!should_suppress_show_mem())
2202 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2203 unsigned long did_some_progress,
2204 unsigned long pages_reclaimed)
2206 /* Do not loop if specifically requested */
2207 if (gfp_mask & __GFP_NORETRY)
2210 /* Always retry if specifically requested */
2211 if (gfp_mask & __GFP_NOFAIL)
2215 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2216 * making forward progress without invoking OOM. Suspend also disables
2217 * storage devices so kswapd will not help. Bail if we are suspending.
2219 if (!did_some_progress && pm_suspended_storage())
2223 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2224 * means __GFP_NOFAIL, but that may not be true in other
2227 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2231 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2232 * specified, then we retry until we no longer reclaim any pages
2233 * (above), or we've reclaimed an order of pages at least as
2234 * large as the allocation's order. In both cases, if the
2235 * allocation still fails, we stop retrying.
2237 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2243 static inline struct page *
2244 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2245 struct zonelist *zonelist, enum zone_type high_zoneidx,
2246 nodemask_t *nodemask, struct zone *preferred_zone,
2247 int classzone_idx, int migratetype)
2251 /* Acquire the per-zone oom lock for each zone */
2252 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2253 schedule_timeout_uninterruptible(1);
2258 * PM-freezer should be notified that there might be an OOM killer on
2259 * its way to kill and wake somebody up. This is too early and we might
2260 * end up not killing anything but false positives are acceptable.
2261 * See freeze_processes.
2266 * Go through the zonelist yet one more time, keep very high watermark
2267 * here, this is only to catch a parallel oom killing, we must fail if
2268 * we're still under heavy pressure.
2270 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2271 order, zonelist, high_zoneidx,
2272 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2273 preferred_zone, classzone_idx, migratetype);
2277 if (!(gfp_mask & __GFP_NOFAIL)) {
2278 /* The OOM killer will not help higher order allocs */
2279 if (order > PAGE_ALLOC_COSTLY_ORDER)
2281 /* The OOM killer does not needlessly kill tasks for lowmem */
2282 if (high_zoneidx < ZONE_NORMAL)
2285 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2286 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2287 * The caller should handle page allocation failure by itself if
2288 * it specifies __GFP_THISNODE.
2289 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2291 if (gfp_mask & __GFP_THISNODE)
2294 /* Exhausted what can be done so it's blamo time */
2295 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2298 oom_zonelist_unlock(zonelist, gfp_mask);
2302 #ifdef CONFIG_COMPACTION
2303 /* Try memory compaction for high-order allocations before reclaim */
2304 static struct page *
2305 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2306 struct zonelist *zonelist, enum zone_type high_zoneidx,
2307 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2308 int classzone_idx, int migratetype, enum migrate_mode mode,
2309 int *contended_compaction, bool *deferred_compaction)
2311 struct zone *last_compact_zone = NULL;
2312 unsigned long compact_result;
2318 current->flags |= PF_MEMALLOC;
2319 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2321 contended_compaction,
2322 &last_compact_zone);
2323 current->flags &= ~PF_MEMALLOC;
2325 switch (compact_result) {
2326 case COMPACT_DEFERRED:
2327 *deferred_compaction = true;
2329 case COMPACT_SKIPPED:
2336 * At least in one zone compaction wasn't deferred or skipped, so let's
2337 * count a compaction stall
2339 count_vm_event(COMPACTSTALL);
2341 /* Page migration frees to the PCP lists but we want merging */
2342 drain_pages(get_cpu());
2345 page = get_page_from_freelist(gfp_mask, nodemask,
2346 order, zonelist, high_zoneidx,
2347 alloc_flags & ~ALLOC_NO_WATERMARKS,
2348 preferred_zone, classzone_idx, migratetype);
2351 struct zone *zone = page_zone(page);
2353 zone->compact_blockskip_flush = false;
2354 compaction_defer_reset(zone, order, true);
2355 count_vm_event(COMPACTSUCCESS);
2360 * last_compact_zone is where try_to_compact_pages thought allocation
2361 * should succeed, so it did not defer compaction. But here we know
2362 * that it didn't succeed, so we do the defer.
2364 if (last_compact_zone && mode != MIGRATE_ASYNC)
2365 defer_compaction(last_compact_zone, order);
2368 * It's bad if compaction run occurs and fails. The most likely reason
2369 * is that pages exist, but not enough to satisfy watermarks.
2371 count_vm_event(COMPACTFAIL);
2378 static inline struct page *
2379 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2380 struct zonelist *zonelist, enum zone_type high_zoneidx,
2381 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2382 int classzone_idx, int migratetype, enum migrate_mode mode,
2383 int *contended_compaction, bool *deferred_compaction)
2387 #endif /* CONFIG_COMPACTION */
2389 /* Perform direct synchronous page reclaim */
2391 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2392 nodemask_t *nodemask)
2394 struct reclaim_state reclaim_state;
2399 /* We now go into synchronous reclaim */
2400 cpuset_memory_pressure_bump();
2401 current->flags |= PF_MEMALLOC;
2402 lockdep_set_current_reclaim_state(gfp_mask);
2403 reclaim_state.reclaimed_slab = 0;
2404 current->reclaim_state = &reclaim_state;
2406 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2408 current->reclaim_state = NULL;
2409 lockdep_clear_current_reclaim_state();
2410 current->flags &= ~PF_MEMALLOC;
2417 /* The really slow allocator path where we enter direct reclaim */
2418 static inline struct page *
2419 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2420 struct zonelist *zonelist, enum zone_type high_zoneidx,
2421 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2422 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2424 struct page *page = NULL;
2425 bool drained = false;
2427 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2429 if (unlikely(!(*did_some_progress)))
2432 /* After successful reclaim, reconsider all zones for allocation */
2433 if (IS_ENABLED(CONFIG_NUMA))
2434 zlc_clear_zones_full(zonelist);
2437 page = get_page_from_freelist(gfp_mask, nodemask, order,
2438 zonelist, high_zoneidx,
2439 alloc_flags & ~ALLOC_NO_WATERMARKS,
2440 preferred_zone, classzone_idx,
2444 * If an allocation failed after direct reclaim, it could be because
2445 * pages are pinned on the per-cpu lists. Drain them and try again
2447 if (!page && !drained) {
2457 * This is called in the allocator slow-path if the allocation request is of
2458 * sufficient urgency to ignore watermarks and take other desperate measures
2460 static inline struct page *
2461 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2462 struct zonelist *zonelist, enum zone_type high_zoneidx,
2463 nodemask_t *nodemask, struct zone *preferred_zone,
2464 int classzone_idx, int migratetype)
2469 page = get_page_from_freelist(gfp_mask, nodemask, order,
2470 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2471 preferred_zone, classzone_idx, migratetype);
2473 if (!page && gfp_mask & __GFP_NOFAIL)
2474 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2475 } while (!page && (gfp_mask & __GFP_NOFAIL));
2480 static void wake_all_kswapds(unsigned int order,
2481 struct zonelist *zonelist,
2482 enum zone_type high_zoneidx,
2483 struct zone *preferred_zone,
2484 nodemask_t *nodemask)
2489 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2490 high_zoneidx, nodemask)
2491 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2495 gfp_to_alloc_flags(gfp_t gfp_mask)
2497 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2498 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2500 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2501 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2504 * The caller may dip into page reserves a bit more if the caller
2505 * cannot run direct reclaim, or if the caller has realtime scheduling
2506 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2507 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2509 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2513 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2514 * if it can't schedule.
2516 if (!(gfp_mask & __GFP_NOMEMALLOC))
2517 alloc_flags |= ALLOC_HARDER;
2519 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2520 * comment for __cpuset_node_allowed_softwall().
2522 alloc_flags &= ~ALLOC_CPUSET;
2523 } else if (unlikely(rt_task(current)) && !in_interrupt())
2524 alloc_flags |= ALLOC_HARDER;
2526 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2527 if (gfp_mask & __GFP_MEMALLOC)
2528 alloc_flags |= ALLOC_NO_WATERMARKS;
2529 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2530 alloc_flags |= ALLOC_NO_WATERMARKS;
2531 else if (!in_interrupt() &&
2532 ((current->flags & PF_MEMALLOC) ||
2533 unlikely(test_thread_flag(TIF_MEMDIE))))
2534 alloc_flags |= ALLOC_NO_WATERMARKS;
2537 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2538 alloc_flags |= ALLOC_CMA;
2543 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2545 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2548 static inline struct page *
2549 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2550 struct zonelist *zonelist, enum zone_type high_zoneidx,
2551 nodemask_t *nodemask, struct zone *preferred_zone,
2552 int classzone_idx, int migratetype)
2554 const gfp_t wait = gfp_mask & __GFP_WAIT;
2555 struct page *page = NULL;
2557 unsigned long pages_reclaimed = 0;
2558 unsigned long did_some_progress;
2559 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2560 bool deferred_compaction = false;
2561 int contended_compaction = COMPACT_CONTENDED_NONE;
2564 * In the slowpath, we sanity check order to avoid ever trying to
2565 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2566 * be using allocators in order of preference for an area that is
2569 if (order >= MAX_ORDER) {
2570 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2575 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2576 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2577 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2578 * using a larger set of nodes after it has established that the
2579 * allowed per node queues are empty and that nodes are
2582 if (IS_ENABLED(CONFIG_NUMA) &&
2583 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2587 if (!(gfp_mask & __GFP_NO_KSWAPD))
2588 wake_all_kswapds(order, zonelist, high_zoneidx,
2589 preferred_zone, nodemask);
2592 * OK, we're below the kswapd watermark and have kicked background
2593 * reclaim. Now things get more complex, so set up alloc_flags according
2594 * to how we want to proceed.
2596 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2599 * Find the true preferred zone if the allocation is unconstrained by
2602 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2603 struct zoneref *preferred_zoneref;
2604 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2605 NULL, &preferred_zone);
2606 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2610 /* This is the last chance, in general, before the goto nopage. */
2611 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2612 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2613 preferred_zone, classzone_idx, migratetype);
2617 /* Allocate without watermarks if the context allows */
2618 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2620 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2621 * the allocation is high priority and these type of
2622 * allocations are system rather than user orientated
2624 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2626 page = __alloc_pages_high_priority(gfp_mask, order,
2627 zonelist, high_zoneidx, nodemask,
2628 preferred_zone, classzone_idx, migratetype);
2634 /* Atomic allocations - we can't balance anything */
2637 * All existing users of the deprecated __GFP_NOFAIL are
2638 * blockable, so warn of any new users that actually allow this
2639 * type of allocation to fail.
2641 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2645 /* Avoid recursion of direct reclaim */
2646 if (current->flags & PF_MEMALLOC)
2649 /* Avoid allocations with no watermarks from looping endlessly */
2650 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2654 * Try direct compaction. The first pass is asynchronous. Subsequent
2655 * attempts after direct reclaim are synchronous
2657 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2658 high_zoneidx, nodemask, alloc_flags,
2660 classzone_idx, migratetype,
2661 migration_mode, &contended_compaction,
2662 &deferred_compaction);
2666 /* Checks for THP-specific high-order allocations */
2667 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2669 * If compaction is deferred for high-order allocations, it is
2670 * because sync compaction recently failed. If this is the case
2671 * and the caller requested a THP allocation, we do not want
2672 * to heavily disrupt the system, so we fail the allocation
2673 * instead of entering direct reclaim.
2675 if (deferred_compaction)
2679 * In all zones where compaction was attempted (and not
2680 * deferred or skipped), lock contention has been detected.
2681 * For THP allocation we do not want to disrupt the others
2682 * so we fallback to base pages instead.
2684 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2688 * If compaction was aborted due to need_resched(), we do not
2689 * want to further increase allocation latency, unless it is
2690 * khugepaged trying to collapse.
2692 if (contended_compaction == COMPACT_CONTENDED_SCHED
2693 && !(current->flags & PF_KTHREAD))
2698 * It can become very expensive to allocate transparent hugepages at
2699 * fault, so use asynchronous memory compaction for THP unless it is
2700 * khugepaged trying to collapse.
2702 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2703 (current->flags & PF_KTHREAD))
2704 migration_mode = MIGRATE_SYNC_LIGHT;
2706 /* Try direct reclaim and then allocating */
2707 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2708 zonelist, high_zoneidx,
2710 alloc_flags, preferred_zone,
2711 classzone_idx, migratetype,
2712 &did_some_progress);
2717 * If we failed to make any progress reclaiming, then we are
2718 * running out of options and have to consider going OOM
2720 if (!did_some_progress) {
2721 if (oom_gfp_allowed(gfp_mask)) {
2722 if (oom_killer_disabled)
2724 /* Coredumps can quickly deplete all memory reserves */
2725 if ((current->flags & PF_DUMPCORE) &&
2726 !(gfp_mask & __GFP_NOFAIL))
2728 page = __alloc_pages_may_oom(gfp_mask, order,
2729 zonelist, high_zoneidx,
2730 nodemask, preferred_zone,
2731 classzone_idx, migratetype);
2735 if (!(gfp_mask & __GFP_NOFAIL)) {
2737 * The oom killer is not called for high-order
2738 * allocations that may fail, so if no progress
2739 * is being made, there are no other options and
2740 * retrying is unlikely to help.
2742 if (order > PAGE_ALLOC_COSTLY_ORDER)
2745 * The oom killer is not called for lowmem
2746 * allocations to prevent needlessly killing
2749 if (high_zoneidx < ZONE_NORMAL)
2757 /* Check if we should retry the allocation */
2758 pages_reclaimed += did_some_progress;
2759 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2761 /* Wait for some write requests to complete then retry */
2762 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2766 * High-order allocations do not necessarily loop after
2767 * direct reclaim and reclaim/compaction depends on compaction
2768 * being called after reclaim so call directly if necessary
2770 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2771 high_zoneidx, nodemask, alloc_flags,
2773 classzone_idx, migratetype,
2774 migration_mode, &contended_compaction,
2775 &deferred_compaction);
2781 warn_alloc_failed(gfp_mask, order, NULL);
2784 if (kmemcheck_enabled)
2785 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2791 * This is the 'heart' of the zoned buddy allocator.
2794 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2795 struct zonelist *zonelist, nodemask_t *nodemask)
2797 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2798 struct zone *preferred_zone;
2799 struct zoneref *preferred_zoneref;
2800 struct page *page = NULL;
2801 int migratetype = gfpflags_to_migratetype(gfp_mask);
2802 unsigned int cpuset_mems_cookie;
2803 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2806 gfp_mask &= gfp_allowed_mask;
2808 lockdep_trace_alloc(gfp_mask);
2810 might_sleep_if(gfp_mask & __GFP_WAIT);
2812 if (should_fail_alloc_page(gfp_mask, order))
2816 * Check the zones suitable for the gfp_mask contain at least one
2817 * valid zone. It's possible to have an empty zonelist as a result
2818 * of GFP_THISNODE and a memoryless node
2820 if (unlikely(!zonelist->_zonerefs->zone))
2823 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2824 alloc_flags |= ALLOC_CMA;
2827 cpuset_mems_cookie = read_mems_allowed_begin();
2829 /* The preferred zone is used for statistics later */
2830 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2831 nodemask ? : &cpuset_current_mems_allowed,
2833 if (!preferred_zone)
2835 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2837 /* First allocation attempt */
2838 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2839 zonelist, high_zoneidx, alloc_flags,
2840 preferred_zone, classzone_idx, migratetype);
2841 if (unlikely(!page)) {
2843 * Runtime PM, block IO and its error handling path
2844 * can deadlock because I/O on the device might not
2847 gfp_mask = memalloc_noio_flags(gfp_mask);
2848 page = __alloc_pages_slowpath(gfp_mask, order,
2849 zonelist, high_zoneidx, nodemask,
2850 preferred_zone, classzone_idx, migratetype);
2853 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2857 * When updating a task's mems_allowed, it is possible to race with
2858 * parallel threads in such a way that an allocation can fail while
2859 * the mask is being updated. If a page allocation is about to fail,
2860 * check if the cpuset changed during allocation and if so, retry.
2862 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2867 EXPORT_SYMBOL(__alloc_pages_nodemask);
2870 * Common helper functions.
2872 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2877 * __get_free_pages() returns a 32-bit address, which cannot represent
2880 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2882 page = alloc_pages(gfp_mask, order);
2885 return (unsigned long) page_address(page);
2887 EXPORT_SYMBOL(__get_free_pages);
2889 unsigned long get_zeroed_page(gfp_t gfp_mask)
2891 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2893 EXPORT_SYMBOL(get_zeroed_page);
2895 void __free_pages(struct page *page, unsigned int order)
2897 if (put_page_testzero(page)) {
2899 free_hot_cold_page(page, false);
2901 __free_pages_ok(page, order);
2905 EXPORT_SYMBOL(__free_pages);
2907 void free_pages(unsigned long addr, unsigned int order)
2910 VM_BUG_ON(!virt_addr_valid((void *)addr));
2911 __free_pages(virt_to_page((void *)addr), order);
2915 EXPORT_SYMBOL(free_pages);
2918 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2919 * of the current memory cgroup.
2921 * It should be used when the caller would like to use kmalloc, but since the
2922 * allocation is large, it has to fall back to the page allocator.
2924 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2927 struct mem_cgroup *memcg = NULL;
2929 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2931 page = alloc_pages(gfp_mask, order);
2932 memcg_kmem_commit_charge(page, memcg, order);
2936 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2939 struct mem_cgroup *memcg = NULL;
2941 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2943 page = alloc_pages_node(nid, gfp_mask, order);
2944 memcg_kmem_commit_charge(page, memcg, order);
2949 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2952 void __free_kmem_pages(struct page *page, unsigned int order)
2954 memcg_kmem_uncharge_pages(page, order);
2955 __free_pages(page, order);
2958 void free_kmem_pages(unsigned long addr, unsigned int order)
2961 VM_BUG_ON(!virt_addr_valid((void *)addr));
2962 __free_kmem_pages(virt_to_page((void *)addr), order);
2966 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2969 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2970 unsigned long used = addr + PAGE_ALIGN(size);
2972 split_page(virt_to_page((void *)addr), order);
2973 while (used < alloc_end) {
2978 return (void *)addr;
2982 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2983 * @size: the number of bytes to allocate
2984 * @gfp_mask: GFP flags for the allocation
2986 * This function is similar to alloc_pages(), except that it allocates the
2987 * minimum number of pages to satisfy the request. alloc_pages() can only
2988 * allocate memory in power-of-two pages.
2990 * This function is also limited by MAX_ORDER.
2992 * Memory allocated by this function must be released by free_pages_exact().
2994 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2996 unsigned int order = get_order(size);
2999 addr = __get_free_pages(gfp_mask, order);
3000 return make_alloc_exact(addr, order, size);
3002 EXPORT_SYMBOL(alloc_pages_exact);
3005 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3007 * @nid: the preferred node ID where memory should be allocated
3008 * @size: the number of bytes to allocate
3009 * @gfp_mask: GFP flags for the allocation
3011 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3013 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3016 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3018 unsigned order = get_order(size);
3019 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3022 return make_alloc_exact((unsigned long)page_address(p), order, size);
3026 * free_pages_exact - release memory allocated via alloc_pages_exact()
3027 * @virt: the value returned by alloc_pages_exact.
3028 * @size: size of allocation, same value as passed to alloc_pages_exact().
3030 * Release the memory allocated by a previous call to alloc_pages_exact.
3032 void free_pages_exact(void *virt, size_t size)
3034 unsigned long addr = (unsigned long)virt;
3035 unsigned long end = addr + PAGE_ALIGN(size);
3037 while (addr < end) {
3042 EXPORT_SYMBOL(free_pages_exact);
3045 * nr_free_zone_pages - count number of pages beyond high watermark
3046 * @offset: The zone index of the highest zone
3048 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3049 * high watermark within all zones at or below a given zone index. For each
3050 * zone, the number of pages is calculated as:
3051 * managed_pages - high_pages
3053 static unsigned long nr_free_zone_pages(int offset)
3058 /* Just pick one node, since fallback list is circular */
3059 unsigned long sum = 0;
3061 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3063 for_each_zone_zonelist(zone, z, zonelist, offset) {
3064 unsigned long size = zone->managed_pages;
3065 unsigned long high = high_wmark_pages(zone);
3074 * nr_free_buffer_pages - count number of pages beyond high watermark
3076 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3077 * watermark within ZONE_DMA and ZONE_NORMAL.
3079 unsigned long nr_free_buffer_pages(void)
3081 return nr_free_zone_pages(gfp_zone(GFP_USER));
3083 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3086 * nr_free_pagecache_pages - count number of pages beyond high watermark
3088 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3089 * high watermark within all zones.
3091 unsigned long nr_free_pagecache_pages(void)
3093 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3096 static inline void show_node(struct zone *zone)
3098 if (IS_ENABLED(CONFIG_NUMA))
3099 printk("Node %d ", zone_to_nid(zone));
3102 void si_meminfo(struct sysinfo *val)
3104 val->totalram = totalram_pages;
3105 val->sharedram = global_page_state(NR_SHMEM);
3106 val->freeram = global_page_state(NR_FREE_PAGES);
3107 val->bufferram = nr_blockdev_pages();
3108 val->totalhigh = totalhigh_pages;
3109 val->freehigh = nr_free_highpages();
3110 val->mem_unit = PAGE_SIZE;
3113 EXPORT_SYMBOL(si_meminfo);
3116 void si_meminfo_node(struct sysinfo *val, int nid)
3118 int zone_type; /* needs to be signed */
3119 unsigned long managed_pages = 0;
3120 pg_data_t *pgdat = NODE_DATA(nid);
3122 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3123 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3124 val->totalram = managed_pages;
3125 val->sharedram = node_page_state(nid, NR_SHMEM);
3126 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3127 #ifdef CONFIG_HIGHMEM
3128 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3129 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3135 val->mem_unit = PAGE_SIZE;
3140 * Determine whether the node should be displayed or not, depending on whether
3141 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3143 bool skip_free_areas_node(unsigned int flags, int nid)
3146 unsigned int cpuset_mems_cookie;
3148 if (!(flags & SHOW_MEM_FILTER_NODES))
3152 cpuset_mems_cookie = read_mems_allowed_begin();
3153 ret = !node_isset(nid, cpuset_current_mems_allowed);
3154 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3159 #define K(x) ((x) << (PAGE_SHIFT-10))
3161 static void show_migration_types(unsigned char type)
3163 static const char types[MIGRATE_TYPES] = {
3164 [MIGRATE_UNMOVABLE] = 'U',
3165 [MIGRATE_RECLAIMABLE] = 'E',
3166 [MIGRATE_MOVABLE] = 'M',
3167 [MIGRATE_RESERVE] = 'R',
3169 [MIGRATE_CMA] = 'C',
3171 #ifdef CONFIG_MEMORY_ISOLATION
3172 [MIGRATE_ISOLATE] = 'I',
3175 char tmp[MIGRATE_TYPES + 1];
3179 for (i = 0; i < MIGRATE_TYPES; i++) {
3180 if (type & (1 << i))
3185 printk("(%s) ", tmp);
3189 * Show free area list (used inside shift_scroll-lock stuff)
3190 * We also calculate the percentage fragmentation. We do this by counting the
3191 * memory on each free list with the exception of the first item on the list.
3192 * Suppresses nodes that are not allowed by current's cpuset if
3193 * SHOW_MEM_FILTER_NODES is passed.
3195 void show_free_areas(unsigned int filter)
3200 for_each_populated_zone(zone) {
3201 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3204 printk("%s per-cpu:\n", zone->name);
3206 for_each_online_cpu(cpu) {
3207 struct per_cpu_pageset *pageset;
3209 pageset = per_cpu_ptr(zone->pageset, cpu);
3211 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3212 cpu, pageset->pcp.high,
3213 pageset->pcp.batch, pageset->pcp.count);
3217 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3218 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3220 " dirty:%lu writeback:%lu unstable:%lu\n"
3221 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3222 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3224 global_page_state(NR_ACTIVE_ANON),
3225 global_page_state(NR_INACTIVE_ANON),
3226 global_page_state(NR_ISOLATED_ANON),
3227 global_page_state(NR_ACTIVE_FILE),
3228 global_page_state(NR_INACTIVE_FILE),
3229 global_page_state(NR_ISOLATED_FILE),
3230 global_page_state(NR_UNEVICTABLE),
3231 global_page_state(NR_FILE_DIRTY),
3232 global_page_state(NR_WRITEBACK),
3233 global_page_state(NR_UNSTABLE_NFS),
3234 global_page_state(NR_FREE_PAGES),
3235 global_page_state(NR_SLAB_RECLAIMABLE),
3236 global_page_state(NR_SLAB_UNRECLAIMABLE),
3237 global_page_state(NR_FILE_MAPPED),
3238 global_page_state(NR_SHMEM),
3239 global_page_state(NR_PAGETABLE),
3240 global_page_state(NR_BOUNCE),
3241 global_page_state(NR_FREE_CMA_PAGES));
3243 for_each_populated_zone(zone) {
3246 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3254 " active_anon:%lukB"
3255 " inactive_anon:%lukB"
3256 " active_file:%lukB"
3257 " inactive_file:%lukB"
3258 " unevictable:%lukB"
3259 " isolated(anon):%lukB"
3260 " isolated(file):%lukB"
3268 " slab_reclaimable:%lukB"
3269 " slab_unreclaimable:%lukB"
3270 " kernel_stack:%lukB"
3275 " writeback_tmp:%lukB"
3276 " pages_scanned:%lu"
3277 " all_unreclaimable? %s"
3280 K(zone_page_state(zone, NR_FREE_PAGES)),
3281 K(min_wmark_pages(zone)),
3282 K(low_wmark_pages(zone)),
3283 K(high_wmark_pages(zone)),
3284 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3285 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3286 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3287 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3288 K(zone_page_state(zone, NR_UNEVICTABLE)),
3289 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3290 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3291 K(zone->present_pages),
3292 K(zone->managed_pages),
3293 K(zone_page_state(zone, NR_MLOCK)),
3294 K(zone_page_state(zone, NR_FILE_DIRTY)),
3295 K(zone_page_state(zone, NR_WRITEBACK)),
3296 K(zone_page_state(zone, NR_FILE_MAPPED)),
3297 K(zone_page_state(zone, NR_SHMEM)),
3298 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3299 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3300 zone_page_state(zone, NR_KERNEL_STACK) *
3302 K(zone_page_state(zone, NR_PAGETABLE)),
3303 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3304 K(zone_page_state(zone, NR_BOUNCE)),
3305 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3306 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3307 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3308 (!zone_reclaimable(zone) ? "yes" : "no")
3310 printk("lowmem_reserve[]:");
3311 for (i = 0; i < MAX_NR_ZONES; i++)
3312 printk(" %ld", zone->lowmem_reserve[i]);
3316 for_each_populated_zone(zone) {
3317 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3318 unsigned char types[MAX_ORDER];
3320 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3323 printk("%s: ", zone->name);
3325 spin_lock_irqsave(&zone->lock, flags);
3326 for (order = 0; order < MAX_ORDER; order++) {
3327 struct free_area *area = &zone->free_area[order];
3330 nr[order] = area->nr_free;
3331 total += nr[order] << order;
3334 for (type = 0; type < MIGRATE_TYPES; type++) {
3335 if (!list_empty(&area->free_list[type]))
3336 types[order] |= 1 << type;
3339 spin_unlock_irqrestore(&zone->lock, flags);
3340 for (order = 0; order < MAX_ORDER; order++) {
3341 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3343 show_migration_types(types[order]);
3345 printk("= %lukB\n", K(total));
3348 hugetlb_show_meminfo();
3350 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3352 show_swap_cache_info();
3355 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3357 zoneref->zone = zone;
3358 zoneref->zone_idx = zone_idx(zone);
3362 * Builds allocation fallback zone lists.
3364 * Add all populated zones of a node to the zonelist.
3366 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3370 enum zone_type zone_type = MAX_NR_ZONES;
3374 zone = pgdat->node_zones + zone_type;
3375 if (populated_zone(zone)) {
3376 zoneref_set_zone(zone,
3377 &zonelist->_zonerefs[nr_zones++]);
3378 check_highest_zone(zone_type);
3380 } while (zone_type);
3388 * 0 = automatic detection of better ordering.
3389 * 1 = order by ([node] distance, -zonetype)
3390 * 2 = order by (-zonetype, [node] distance)
3392 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3393 * the same zonelist. So only NUMA can configure this param.
3395 #define ZONELIST_ORDER_DEFAULT 0
3396 #define ZONELIST_ORDER_NODE 1
3397 #define ZONELIST_ORDER_ZONE 2
3399 /* zonelist order in the kernel.
3400 * set_zonelist_order() will set this to NODE or ZONE.
3402 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3403 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3407 /* The value user specified ....changed by config */
3408 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3409 /* string for sysctl */
3410 #define NUMA_ZONELIST_ORDER_LEN 16
3411 char numa_zonelist_order[16] = "default";
3414 * interface for configure zonelist ordering.
3415 * command line option "numa_zonelist_order"
3416 * = "[dD]efault - default, automatic configuration.
3417 * = "[nN]ode - order by node locality, then by zone within node
3418 * = "[zZ]one - order by zone, then by locality within zone
3421 static int __parse_numa_zonelist_order(char *s)
3423 if (*s == 'd' || *s == 'D') {
3424 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3425 } else if (*s == 'n' || *s == 'N') {
3426 user_zonelist_order = ZONELIST_ORDER_NODE;
3427 } else if (*s == 'z' || *s == 'Z') {
3428 user_zonelist_order = ZONELIST_ORDER_ZONE;
3431 "Ignoring invalid numa_zonelist_order value: "
3438 static __init int setup_numa_zonelist_order(char *s)
3445 ret = __parse_numa_zonelist_order(s);
3447 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3451 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3454 * sysctl handler for numa_zonelist_order
3456 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3457 void __user *buffer, size_t *length,
3460 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3462 static DEFINE_MUTEX(zl_order_mutex);
3464 mutex_lock(&zl_order_mutex);
3466 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3470 strcpy(saved_string, (char *)table->data);
3472 ret = proc_dostring(table, write, buffer, length, ppos);
3476 int oldval = user_zonelist_order;
3478 ret = __parse_numa_zonelist_order((char *)table->data);
3481 * bogus value. restore saved string
3483 strncpy((char *)table->data, saved_string,
3484 NUMA_ZONELIST_ORDER_LEN);
3485 user_zonelist_order = oldval;
3486 } else if (oldval != user_zonelist_order) {
3487 mutex_lock(&zonelists_mutex);
3488 build_all_zonelists(NULL, NULL);
3489 mutex_unlock(&zonelists_mutex);
3493 mutex_unlock(&zl_order_mutex);
3498 #define MAX_NODE_LOAD (nr_online_nodes)
3499 static int node_load[MAX_NUMNODES];
3502 * find_next_best_node - find the next node that should appear in a given node's fallback list
3503 * @node: node whose fallback list we're appending
3504 * @used_node_mask: nodemask_t of already used nodes
3506 * We use a number of factors to determine which is the next node that should
3507 * appear on a given node's fallback list. The node should not have appeared
3508 * already in @node's fallback list, and it should be the next closest node
3509 * according to the distance array (which contains arbitrary distance values
3510 * from each node to each node in the system), and should also prefer nodes
3511 * with no CPUs, since presumably they'll have very little allocation pressure
3512 * on them otherwise.
3513 * It returns -1 if no node is found.
3515 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3518 int min_val = INT_MAX;
3519 int best_node = NUMA_NO_NODE;
3520 const struct cpumask *tmp = cpumask_of_node(0);
3522 /* Use the local node if we haven't already */
3523 if (!node_isset(node, *used_node_mask)) {
3524 node_set(node, *used_node_mask);
3528 for_each_node_state(n, N_MEMORY) {
3530 /* Don't want a node to appear more than once */
3531 if (node_isset(n, *used_node_mask))
3534 /* Use the distance array to find the distance */
3535 val = node_distance(node, n);
3537 /* Penalize nodes under us ("prefer the next node") */
3540 /* Give preference to headless and unused nodes */
3541 tmp = cpumask_of_node(n);
3542 if (!cpumask_empty(tmp))
3543 val += PENALTY_FOR_NODE_WITH_CPUS;
3545 /* Slight preference for less loaded node */
3546 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3547 val += node_load[n];
3549 if (val < min_val) {
3556 node_set(best_node, *used_node_mask);
3563 * Build zonelists ordered by node and zones within node.
3564 * This results in maximum locality--normal zone overflows into local
3565 * DMA zone, if any--but risks exhausting DMA zone.
3567 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3570 struct zonelist *zonelist;
3572 zonelist = &pgdat->node_zonelists[0];
3573 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3575 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3576 zonelist->_zonerefs[j].zone = NULL;
3577 zonelist->_zonerefs[j].zone_idx = 0;
3581 * Build gfp_thisnode zonelists
3583 static void build_thisnode_zonelists(pg_data_t *pgdat)
3586 struct zonelist *zonelist;
3588 zonelist = &pgdat->node_zonelists[1];
3589 j = build_zonelists_node(pgdat, zonelist, 0);
3590 zonelist->_zonerefs[j].zone = NULL;
3591 zonelist->_zonerefs[j].zone_idx = 0;
3595 * Build zonelists ordered by zone and nodes within zones.
3596 * This results in conserving DMA zone[s] until all Normal memory is
3597 * exhausted, but results in overflowing to remote node while memory
3598 * may still exist in local DMA zone.
3600 static int node_order[MAX_NUMNODES];
3602 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3605 int zone_type; /* needs to be signed */
3607 struct zonelist *zonelist;
3609 zonelist = &pgdat->node_zonelists[0];
3611 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3612 for (j = 0; j < nr_nodes; j++) {
3613 node = node_order[j];
3614 z = &NODE_DATA(node)->node_zones[zone_type];
3615 if (populated_zone(z)) {
3617 &zonelist->_zonerefs[pos++]);
3618 check_highest_zone(zone_type);
3622 zonelist->_zonerefs[pos].zone = NULL;
3623 zonelist->_zonerefs[pos].zone_idx = 0;
3626 #if defined(CONFIG_64BIT)
3628 * Devices that require DMA32/DMA are relatively rare and do not justify a
3629 * penalty to every machine in case the specialised case applies. Default
3630 * to Node-ordering on 64-bit NUMA machines
3632 static int default_zonelist_order(void)
3634 return ZONELIST_ORDER_NODE;
3638 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3639 * by the kernel. If processes running on node 0 deplete the low memory zone
3640 * then reclaim will occur more frequency increasing stalls and potentially
3641 * be easier to OOM if a large percentage of the zone is under writeback or
3642 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3643 * Hence, default to zone ordering on 32-bit.
3645 static int default_zonelist_order(void)
3647 return ZONELIST_ORDER_ZONE;
3649 #endif /* CONFIG_64BIT */
3651 static void set_zonelist_order(void)
3653 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3654 current_zonelist_order = default_zonelist_order();
3656 current_zonelist_order = user_zonelist_order;
3659 static void build_zonelists(pg_data_t *pgdat)
3663 nodemask_t used_mask;
3664 int local_node, prev_node;
3665 struct zonelist *zonelist;
3666 int order = current_zonelist_order;
3668 /* initialize zonelists */
3669 for (i = 0; i < MAX_ZONELISTS; i++) {
3670 zonelist = pgdat->node_zonelists + i;
3671 zonelist->_zonerefs[0].zone = NULL;
3672 zonelist->_zonerefs[0].zone_idx = 0;
3675 /* NUMA-aware ordering of nodes */
3676 local_node = pgdat->node_id;
3677 load = nr_online_nodes;
3678 prev_node = local_node;
3679 nodes_clear(used_mask);
3681 memset(node_order, 0, sizeof(node_order));
3684 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3686 * We don't want to pressure a particular node.
3687 * So adding penalty to the first node in same
3688 * distance group to make it round-robin.
3690 if (node_distance(local_node, node) !=
3691 node_distance(local_node, prev_node))
3692 node_load[node] = load;
3696 if (order == ZONELIST_ORDER_NODE)
3697 build_zonelists_in_node_order(pgdat, node);
3699 node_order[j++] = node; /* remember order */
3702 if (order == ZONELIST_ORDER_ZONE) {
3703 /* calculate node order -- i.e., DMA last! */
3704 build_zonelists_in_zone_order(pgdat, j);
3707 build_thisnode_zonelists(pgdat);
3710 /* Construct the zonelist performance cache - see further mmzone.h */
3711 static void build_zonelist_cache(pg_data_t *pgdat)
3713 struct zonelist *zonelist;
3714 struct zonelist_cache *zlc;
3717 zonelist = &pgdat->node_zonelists[0];
3718 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3719 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3720 for (z = zonelist->_zonerefs; z->zone; z++)
3721 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3724 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3726 * Return node id of node used for "local" allocations.
3727 * I.e., first node id of first zone in arg node's generic zonelist.
3728 * Used for initializing percpu 'numa_mem', which is used primarily
3729 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3731 int local_memory_node(int node)
3735 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3736 gfp_zone(GFP_KERNEL),
3743 #else /* CONFIG_NUMA */
3745 static void set_zonelist_order(void)
3747 current_zonelist_order = ZONELIST_ORDER_ZONE;
3750 static void build_zonelists(pg_data_t *pgdat)
3752 int node, local_node;
3754 struct zonelist *zonelist;
3756 local_node = pgdat->node_id;
3758 zonelist = &pgdat->node_zonelists[0];
3759 j = build_zonelists_node(pgdat, zonelist, 0);
3762 * Now we build the zonelist so that it contains the zones
3763 * of all the other nodes.
3764 * We don't want to pressure a particular node, so when
3765 * building the zones for node N, we make sure that the
3766 * zones coming right after the local ones are those from
3767 * node N+1 (modulo N)
3769 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3770 if (!node_online(node))
3772 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3774 for (node = 0; node < local_node; node++) {
3775 if (!node_online(node))
3777 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3780 zonelist->_zonerefs[j].zone = NULL;
3781 zonelist->_zonerefs[j].zone_idx = 0;
3784 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3785 static void build_zonelist_cache(pg_data_t *pgdat)
3787 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3790 #endif /* CONFIG_NUMA */
3793 * Boot pageset table. One per cpu which is going to be used for all
3794 * zones and all nodes. The parameters will be set in such a way
3795 * that an item put on a list will immediately be handed over to
3796 * the buddy list. This is safe since pageset manipulation is done
3797 * with interrupts disabled.
3799 * The boot_pagesets must be kept even after bootup is complete for
3800 * unused processors and/or zones. They do play a role for bootstrapping
3801 * hotplugged processors.
3803 * zoneinfo_show() and maybe other functions do
3804 * not check if the processor is online before following the pageset pointer.
3805 * Other parts of the kernel may not check if the zone is available.
3807 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3808 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3809 static void setup_zone_pageset(struct zone *zone);
3812 * Global mutex to protect against size modification of zonelists
3813 * as well as to serialize pageset setup for the new populated zone.
3815 DEFINE_MUTEX(zonelists_mutex);
3817 /* return values int ....just for stop_machine() */
3818 static int __build_all_zonelists(void *data)
3822 pg_data_t *self = data;
3825 memset(node_load, 0, sizeof(node_load));
3828 if (self && !node_online(self->node_id)) {
3829 build_zonelists(self);
3830 build_zonelist_cache(self);
3833 for_each_online_node(nid) {
3834 pg_data_t *pgdat = NODE_DATA(nid);
3836 build_zonelists(pgdat);
3837 build_zonelist_cache(pgdat);
3841 * Initialize the boot_pagesets that are going to be used
3842 * for bootstrapping processors. The real pagesets for
3843 * each zone will be allocated later when the per cpu
3844 * allocator is available.
3846 * boot_pagesets are used also for bootstrapping offline
3847 * cpus if the system is already booted because the pagesets
3848 * are needed to initialize allocators on a specific cpu too.
3849 * F.e. the percpu allocator needs the page allocator which
3850 * needs the percpu allocator in order to allocate its pagesets
3851 * (a chicken-egg dilemma).
3853 for_each_possible_cpu(cpu) {
3854 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3856 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3858 * We now know the "local memory node" for each node--
3859 * i.e., the node of the first zone in the generic zonelist.
3860 * Set up numa_mem percpu variable for on-line cpus. During
3861 * boot, only the boot cpu should be on-line; we'll init the
3862 * secondary cpus' numa_mem as they come on-line. During
3863 * node/memory hotplug, we'll fixup all on-line cpus.
3865 if (cpu_online(cpu))
3866 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3874 * Called with zonelists_mutex held always
3875 * unless system_state == SYSTEM_BOOTING.
3877 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3879 set_zonelist_order();
3881 if (system_state == SYSTEM_BOOTING) {
3882 __build_all_zonelists(NULL);
3883 mminit_verify_zonelist();
3884 cpuset_init_current_mems_allowed();
3886 #ifdef CONFIG_MEMORY_HOTPLUG
3888 setup_zone_pageset(zone);
3890 /* we have to stop all cpus to guarantee there is no user
3892 stop_machine(__build_all_zonelists, pgdat, NULL);
3893 /* cpuset refresh routine should be here */
3895 vm_total_pages = nr_free_pagecache_pages();
3897 * Disable grouping by mobility if the number of pages in the
3898 * system is too low to allow the mechanism to work. It would be
3899 * more accurate, but expensive to check per-zone. This check is
3900 * made on memory-hotadd so a system can start with mobility
3901 * disabled and enable it later
3903 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3904 page_group_by_mobility_disabled = 1;
3906 page_group_by_mobility_disabled = 0;
3908 printk("Built %i zonelists in %s order, mobility grouping %s. "
3909 "Total pages: %ld\n",
3911 zonelist_order_name[current_zonelist_order],
3912 page_group_by_mobility_disabled ? "off" : "on",
3915 printk("Policy zone: %s\n", zone_names[policy_zone]);
3920 * Helper functions to size the waitqueue hash table.
3921 * Essentially these want to choose hash table sizes sufficiently
3922 * large so that collisions trying to wait on pages are rare.
3923 * But in fact, the number of active page waitqueues on typical
3924 * systems is ridiculously low, less than 200. So this is even
3925 * conservative, even though it seems large.
3927 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3928 * waitqueues, i.e. the size of the waitq table given the number of pages.
3930 #define PAGES_PER_WAITQUEUE 256
3932 #ifndef CONFIG_MEMORY_HOTPLUG
3933 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3935 unsigned long size = 1;
3937 pages /= PAGES_PER_WAITQUEUE;
3939 while (size < pages)
3943 * Once we have dozens or even hundreds of threads sleeping
3944 * on IO we've got bigger problems than wait queue collision.
3945 * Limit the size of the wait table to a reasonable size.
3947 size = min(size, 4096UL);
3949 return max(size, 4UL);
3953 * A zone's size might be changed by hot-add, so it is not possible to determine
3954 * a suitable size for its wait_table. So we use the maximum size now.
3956 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3958 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3959 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3960 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3962 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3963 * or more by the traditional way. (See above). It equals:
3965 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3966 * ia64(16K page size) : = ( 8G + 4M)byte.
3967 * powerpc (64K page size) : = (32G +16M)byte.
3969 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3976 * This is an integer logarithm so that shifts can be used later
3977 * to extract the more random high bits from the multiplicative
3978 * hash function before the remainder is taken.
3980 static inline unsigned long wait_table_bits(unsigned long size)
3986 * Check if a pageblock contains reserved pages
3988 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3992 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3993 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4000 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4001 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4002 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4003 * higher will lead to a bigger reserve which will get freed as contiguous
4004 * blocks as reclaim kicks in
4006 static void setup_zone_migrate_reserve(struct zone *zone)
4008 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4010 unsigned long block_migratetype;
4015 * Get the start pfn, end pfn and the number of blocks to reserve
4016 * We have to be careful to be aligned to pageblock_nr_pages to
4017 * make sure that we always check pfn_valid for the first page in
4020 start_pfn = zone->zone_start_pfn;
4021 end_pfn = zone_end_pfn(zone);
4022 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4023 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4027 * Reserve blocks are generally in place to help high-order atomic
4028 * allocations that are short-lived. A min_free_kbytes value that
4029 * would result in more than 2 reserve blocks for atomic allocations
4030 * is assumed to be in place to help anti-fragmentation for the
4031 * future allocation of hugepages at runtime.
4033 reserve = min(2, reserve);
4034 old_reserve = zone->nr_migrate_reserve_block;
4036 /* When memory hot-add, we almost always need to do nothing */
4037 if (reserve == old_reserve)
4039 zone->nr_migrate_reserve_block = reserve;
4041 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4042 if (!pfn_valid(pfn))
4044 page = pfn_to_page(pfn);
4046 /* Watch out for overlapping nodes */
4047 if (page_to_nid(page) != zone_to_nid(zone))
4050 block_migratetype = get_pageblock_migratetype(page);
4052 /* Only test what is necessary when the reserves are not met */
4055 * Blocks with reserved pages will never free, skip
4058 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4059 if (pageblock_is_reserved(pfn, block_end_pfn))
4062 /* If this block is reserved, account for it */
4063 if (block_migratetype == MIGRATE_RESERVE) {
4068 /* Suitable for reserving if this block is movable */
4069 if (block_migratetype == MIGRATE_MOVABLE) {
4070 set_pageblock_migratetype(page,
4072 move_freepages_block(zone, page,
4077 } else if (!old_reserve) {
4079 * At boot time we don't need to scan the whole zone
4080 * for turning off MIGRATE_RESERVE.
4086 * If the reserve is met and this is a previous reserved block,
4089 if (block_migratetype == MIGRATE_RESERVE) {
4090 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4091 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4097 * Initially all pages are reserved - free ones are freed
4098 * up by free_all_bootmem() once the early boot process is
4099 * done. Non-atomic initialization, single-pass.
4101 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4102 unsigned long start_pfn, enum memmap_context context)
4105 unsigned long end_pfn = start_pfn + size;
4109 if (highest_memmap_pfn < end_pfn - 1)
4110 highest_memmap_pfn = end_pfn - 1;
4112 z = &NODE_DATA(nid)->node_zones[zone];
4113 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4115 * There can be holes in boot-time mem_map[]s
4116 * handed to this function. They do not
4117 * exist on hotplugged memory.
4119 if (context == MEMMAP_EARLY) {
4120 if (!early_pfn_valid(pfn))
4122 if (!early_pfn_in_nid(pfn, nid))
4125 page = pfn_to_page(pfn);
4126 set_page_links(page, zone, nid, pfn);
4127 mminit_verify_page_links(page, zone, nid, pfn);
4128 init_page_count(page);
4129 page_mapcount_reset(page);
4130 page_cpupid_reset_last(page);
4131 SetPageReserved(page);
4133 * Mark the block movable so that blocks are reserved for
4134 * movable at startup. This will force kernel allocations
4135 * to reserve their blocks rather than leaking throughout
4136 * the address space during boot when many long-lived
4137 * kernel allocations are made. Later some blocks near
4138 * the start are marked MIGRATE_RESERVE by
4139 * setup_zone_migrate_reserve()
4141 * bitmap is created for zone's valid pfn range. but memmap
4142 * can be created for invalid pages (for alignment)
4143 * check here not to call set_pageblock_migratetype() against
4146 if ((z->zone_start_pfn <= pfn)
4147 && (pfn < zone_end_pfn(z))
4148 && !(pfn & (pageblock_nr_pages - 1)))
4149 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4151 INIT_LIST_HEAD(&page->lru);
4152 #ifdef WANT_PAGE_VIRTUAL
4153 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4154 if (!is_highmem_idx(zone))
4155 set_page_address(page, __va(pfn << PAGE_SHIFT));
4160 static void __meminit zone_init_free_lists(struct zone *zone)
4162 unsigned int order, t;
4163 for_each_migratetype_order(order, t) {
4164 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4165 zone->free_area[order].nr_free = 0;
4169 #ifndef __HAVE_ARCH_MEMMAP_INIT
4170 #define memmap_init(size, nid, zone, start_pfn) \
4171 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4174 static int zone_batchsize(struct zone *zone)
4180 * The per-cpu-pages pools are set to around 1000th of the
4181 * size of the zone. But no more than 1/2 of a meg.
4183 * OK, so we don't know how big the cache is. So guess.
4185 batch = zone->managed_pages / 1024;
4186 if (batch * PAGE_SIZE > 512 * 1024)
4187 batch = (512 * 1024) / PAGE_SIZE;
4188 batch /= 4; /* We effectively *= 4 below */
4193 * Clamp the batch to a 2^n - 1 value. Having a power
4194 * of 2 value was found to be more likely to have
4195 * suboptimal cache aliasing properties in some cases.
4197 * For example if 2 tasks are alternately allocating
4198 * batches of pages, one task can end up with a lot
4199 * of pages of one half of the possible page colors
4200 * and the other with pages of the other colors.
4202 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4207 /* The deferral and batching of frees should be suppressed under NOMMU
4210 * The problem is that NOMMU needs to be able to allocate large chunks
4211 * of contiguous memory as there's no hardware page translation to
4212 * assemble apparent contiguous memory from discontiguous pages.
4214 * Queueing large contiguous runs of pages for batching, however,
4215 * causes the pages to actually be freed in smaller chunks. As there
4216 * can be a significant delay between the individual batches being
4217 * recycled, this leads to the once large chunks of space being
4218 * fragmented and becoming unavailable for high-order allocations.
4225 * pcp->high and pcp->batch values are related and dependent on one another:
4226 * ->batch must never be higher then ->high.
4227 * The following function updates them in a safe manner without read side
4230 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4231 * those fields changing asynchronously (acording the the above rule).
4233 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4234 * outside of boot time (or some other assurance that no concurrent updaters
4237 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4238 unsigned long batch)
4240 /* start with a fail safe value for batch */
4244 /* Update high, then batch, in order */
4251 /* a companion to pageset_set_high() */
4252 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4254 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4257 static void pageset_init(struct per_cpu_pageset *p)
4259 struct per_cpu_pages *pcp;
4262 memset(p, 0, sizeof(*p));
4266 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4267 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4270 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4273 pageset_set_batch(p, batch);
4277 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4278 * to the value high for the pageset p.
4280 static void pageset_set_high(struct per_cpu_pageset *p,
4283 unsigned long batch = max(1UL, high / 4);
4284 if ((high / 4) > (PAGE_SHIFT * 8))
4285 batch = PAGE_SHIFT * 8;
4287 pageset_update(&p->pcp, high, batch);
4290 static void pageset_set_high_and_batch(struct zone *zone,
4291 struct per_cpu_pageset *pcp)
4293 if (percpu_pagelist_fraction)
4294 pageset_set_high(pcp,
4295 (zone->managed_pages /
4296 percpu_pagelist_fraction));
4298 pageset_set_batch(pcp, zone_batchsize(zone));
4301 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4303 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4306 pageset_set_high_and_batch(zone, pcp);
4309 static void __meminit setup_zone_pageset(struct zone *zone)
4312 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4313 for_each_possible_cpu(cpu)
4314 zone_pageset_init(zone, cpu);
4318 * Allocate per cpu pagesets and initialize them.
4319 * Before this call only boot pagesets were available.
4321 void __init setup_per_cpu_pageset(void)
4325 for_each_populated_zone(zone)
4326 setup_zone_pageset(zone);
4329 static noinline __init_refok
4330 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4336 * The per-page waitqueue mechanism uses hashed waitqueues
4339 zone->wait_table_hash_nr_entries =
4340 wait_table_hash_nr_entries(zone_size_pages);
4341 zone->wait_table_bits =
4342 wait_table_bits(zone->wait_table_hash_nr_entries);
4343 alloc_size = zone->wait_table_hash_nr_entries
4344 * sizeof(wait_queue_head_t);
4346 if (!slab_is_available()) {
4347 zone->wait_table = (wait_queue_head_t *)
4348 memblock_virt_alloc_node_nopanic(
4349 alloc_size, zone->zone_pgdat->node_id);
4352 * This case means that a zone whose size was 0 gets new memory
4353 * via memory hot-add.
4354 * But it may be the case that a new node was hot-added. In
4355 * this case vmalloc() will not be able to use this new node's
4356 * memory - this wait_table must be initialized to use this new
4357 * node itself as well.
4358 * To use this new node's memory, further consideration will be
4361 zone->wait_table = vmalloc(alloc_size);
4363 if (!zone->wait_table)
4366 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4367 init_waitqueue_head(zone->wait_table + i);
4372 static __meminit void zone_pcp_init(struct zone *zone)
4375 * per cpu subsystem is not up at this point. The following code
4376 * relies on the ability of the linker to provide the
4377 * offset of a (static) per cpu variable into the per cpu area.
4379 zone->pageset = &boot_pageset;
4381 if (populated_zone(zone))
4382 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4383 zone->name, zone->present_pages,
4384 zone_batchsize(zone));
4387 int __meminit init_currently_empty_zone(struct zone *zone,
4388 unsigned long zone_start_pfn,
4390 enum memmap_context context)
4392 struct pglist_data *pgdat = zone->zone_pgdat;
4394 ret = zone_wait_table_init(zone, size);
4397 pgdat->nr_zones = zone_idx(zone) + 1;
4399 zone->zone_start_pfn = zone_start_pfn;
4401 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4402 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4404 (unsigned long)zone_idx(zone),
4405 zone_start_pfn, (zone_start_pfn + size));
4407 zone_init_free_lists(zone);
4412 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4413 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4415 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4417 int __meminit __early_pfn_to_nid(unsigned long pfn)
4419 unsigned long start_pfn, end_pfn;
4422 * NOTE: The following SMP-unsafe globals are only used early in boot
4423 * when the kernel is running single-threaded.
4425 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4426 static int __meminitdata last_nid;
4428 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4431 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4433 last_start_pfn = start_pfn;
4434 last_end_pfn = end_pfn;
4440 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4442 int __meminit early_pfn_to_nid(unsigned long pfn)
4446 nid = __early_pfn_to_nid(pfn);
4449 /* just returns 0 */
4453 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4454 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4458 nid = __early_pfn_to_nid(pfn);
4459 if (nid >= 0 && nid != node)
4466 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4467 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4468 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4470 * If an architecture guarantees that all ranges registered contain no holes
4471 * and may be freed, this this function may be used instead of calling
4472 * memblock_free_early_nid() manually.
4474 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4476 unsigned long start_pfn, end_pfn;
4479 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4480 start_pfn = min(start_pfn, max_low_pfn);
4481 end_pfn = min(end_pfn, max_low_pfn);
4483 if (start_pfn < end_pfn)
4484 memblock_free_early_nid(PFN_PHYS(start_pfn),
4485 (end_pfn - start_pfn) << PAGE_SHIFT,
4491 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4492 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4494 * If an architecture guarantees that all ranges registered contain no holes and may
4495 * be freed, this function may be used instead of calling memory_present() manually.
4497 void __init sparse_memory_present_with_active_regions(int nid)
4499 unsigned long start_pfn, end_pfn;
4502 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4503 memory_present(this_nid, start_pfn, end_pfn);
4507 * get_pfn_range_for_nid - Return the start and end page frames for a node
4508 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4509 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4510 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4512 * It returns the start and end page frame of a node based on information
4513 * provided by memblock_set_node(). If called for a node
4514 * with no available memory, a warning is printed and the start and end
4517 void __meminit get_pfn_range_for_nid(unsigned int nid,
4518 unsigned long *start_pfn, unsigned long *end_pfn)
4520 unsigned long this_start_pfn, this_end_pfn;
4526 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4527 *start_pfn = min(*start_pfn, this_start_pfn);
4528 *end_pfn = max(*end_pfn, this_end_pfn);
4531 if (*start_pfn == -1UL)
4536 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4537 * assumption is made that zones within a node are ordered in monotonic
4538 * increasing memory addresses so that the "highest" populated zone is used
4540 static void __init find_usable_zone_for_movable(void)
4543 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4544 if (zone_index == ZONE_MOVABLE)
4547 if (arch_zone_highest_possible_pfn[zone_index] >
4548 arch_zone_lowest_possible_pfn[zone_index])
4552 VM_BUG_ON(zone_index == -1);
4553 movable_zone = zone_index;
4557 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4558 * because it is sized independent of architecture. Unlike the other zones,
4559 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4560 * in each node depending on the size of each node and how evenly kernelcore
4561 * is distributed. This helper function adjusts the zone ranges
4562 * provided by the architecture for a given node by using the end of the
4563 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4564 * zones within a node are in order of monotonic increases memory addresses
4566 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4567 unsigned long zone_type,
4568 unsigned long node_start_pfn,
4569 unsigned long node_end_pfn,
4570 unsigned long *zone_start_pfn,
4571 unsigned long *zone_end_pfn)
4573 /* Only adjust if ZONE_MOVABLE is on this node */
4574 if (zone_movable_pfn[nid]) {
4575 /* Size ZONE_MOVABLE */
4576 if (zone_type == ZONE_MOVABLE) {
4577 *zone_start_pfn = zone_movable_pfn[nid];
4578 *zone_end_pfn = min(node_end_pfn,
4579 arch_zone_highest_possible_pfn[movable_zone]);
4581 /* Adjust for ZONE_MOVABLE starting within this range */
4582 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4583 *zone_end_pfn > zone_movable_pfn[nid]) {
4584 *zone_end_pfn = zone_movable_pfn[nid];
4586 /* Check if this whole range is within ZONE_MOVABLE */
4587 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4588 *zone_start_pfn = *zone_end_pfn;
4593 * Return the number of pages a zone spans in a node, including holes
4594 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4596 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4597 unsigned long zone_type,
4598 unsigned long node_start_pfn,
4599 unsigned long node_end_pfn,
4600 unsigned long *ignored)
4602 unsigned long zone_start_pfn, zone_end_pfn;
4604 /* Get the start and end of the zone */
4605 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4606 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4607 adjust_zone_range_for_zone_movable(nid, zone_type,
4608 node_start_pfn, node_end_pfn,
4609 &zone_start_pfn, &zone_end_pfn);
4611 /* Check that this node has pages within the zone's required range */
4612 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4615 /* Move the zone boundaries inside the node if necessary */
4616 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4617 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4619 /* Return the spanned pages */
4620 return zone_end_pfn - zone_start_pfn;
4624 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4625 * then all holes in the requested range will be accounted for.
4627 unsigned long __meminit __absent_pages_in_range(int nid,
4628 unsigned long range_start_pfn,
4629 unsigned long range_end_pfn)
4631 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4632 unsigned long start_pfn, end_pfn;
4635 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4636 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4637 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4638 nr_absent -= end_pfn - start_pfn;
4644 * absent_pages_in_range - Return number of page frames in holes within a range
4645 * @start_pfn: The start PFN to start searching for holes
4646 * @end_pfn: The end PFN to stop searching for holes
4648 * It returns the number of pages frames in memory holes within a range.
4650 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4651 unsigned long end_pfn)
4653 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4656 /* Return the number of page frames in holes in a zone on a node */
4657 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4658 unsigned long zone_type,
4659 unsigned long node_start_pfn,
4660 unsigned long node_end_pfn,
4661 unsigned long *ignored)
4663 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4664 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4665 unsigned long zone_start_pfn, zone_end_pfn;
4667 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4668 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4670 adjust_zone_range_for_zone_movable(nid, zone_type,
4671 node_start_pfn, node_end_pfn,
4672 &zone_start_pfn, &zone_end_pfn);
4673 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4676 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4677 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4678 unsigned long zone_type,
4679 unsigned long node_start_pfn,
4680 unsigned long node_end_pfn,
4681 unsigned long *zones_size)
4683 return zones_size[zone_type];
4686 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4687 unsigned long zone_type,
4688 unsigned long node_start_pfn,
4689 unsigned long node_end_pfn,
4690 unsigned long *zholes_size)
4695 return zholes_size[zone_type];
4698 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4700 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4701 unsigned long node_start_pfn,
4702 unsigned long node_end_pfn,
4703 unsigned long *zones_size,
4704 unsigned long *zholes_size)
4706 unsigned long realtotalpages, totalpages = 0;
4709 for (i = 0; i < MAX_NR_ZONES; i++)
4710 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4714 pgdat->node_spanned_pages = totalpages;
4716 realtotalpages = totalpages;
4717 for (i = 0; i < MAX_NR_ZONES; i++)
4719 zone_absent_pages_in_node(pgdat->node_id, i,
4720 node_start_pfn, node_end_pfn,
4722 pgdat->node_present_pages = realtotalpages;
4723 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4727 #ifndef CONFIG_SPARSEMEM
4729 * Calculate the size of the zone->blockflags rounded to an unsigned long
4730 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4731 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4732 * round what is now in bits to nearest long in bits, then return it in
4735 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4737 unsigned long usemapsize;
4739 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4740 usemapsize = roundup(zonesize, pageblock_nr_pages);
4741 usemapsize = usemapsize >> pageblock_order;
4742 usemapsize *= NR_PAGEBLOCK_BITS;
4743 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4745 return usemapsize / 8;
4748 static void __init setup_usemap(struct pglist_data *pgdat,
4750 unsigned long zone_start_pfn,
4751 unsigned long zonesize)
4753 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4754 zone->pageblock_flags = NULL;
4756 zone->pageblock_flags =
4757 memblock_virt_alloc_node_nopanic(usemapsize,
4761 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4762 unsigned long zone_start_pfn, unsigned long zonesize) {}
4763 #endif /* CONFIG_SPARSEMEM */
4765 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4767 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4768 void __paginginit set_pageblock_order(void)
4772 /* Check that pageblock_nr_pages has not already been setup */
4773 if (pageblock_order)
4776 if (HPAGE_SHIFT > PAGE_SHIFT)
4777 order = HUGETLB_PAGE_ORDER;
4779 order = MAX_ORDER - 1;
4782 * Assume the largest contiguous order of interest is a huge page.
4783 * This value may be variable depending on boot parameters on IA64 and
4786 pageblock_order = order;
4788 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4791 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4792 * is unused as pageblock_order is set at compile-time. See
4793 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4796 void __paginginit set_pageblock_order(void)
4800 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4802 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4803 unsigned long present_pages)
4805 unsigned long pages = spanned_pages;
4808 * Provide a more accurate estimation if there are holes within
4809 * the zone and SPARSEMEM is in use. If there are holes within the
4810 * zone, each populated memory region may cost us one or two extra
4811 * memmap pages due to alignment because memmap pages for each
4812 * populated regions may not naturally algined on page boundary.
4813 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4815 if (spanned_pages > present_pages + (present_pages >> 4) &&
4816 IS_ENABLED(CONFIG_SPARSEMEM))
4817 pages = present_pages;
4819 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4823 * Set up the zone data structures:
4824 * - mark all pages reserved
4825 * - mark all memory queues empty
4826 * - clear the memory bitmaps
4828 * NOTE: pgdat should get zeroed by caller.
4830 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4831 unsigned long node_start_pfn, unsigned long node_end_pfn,
4832 unsigned long *zones_size, unsigned long *zholes_size)
4835 int nid = pgdat->node_id;
4836 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4839 pgdat_resize_init(pgdat);
4840 #ifdef CONFIG_NUMA_BALANCING
4841 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4842 pgdat->numabalancing_migrate_nr_pages = 0;
4843 pgdat->numabalancing_migrate_next_window = jiffies;
4845 init_waitqueue_head(&pgdat->kswapd_wait);
4846 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4847 pgdat_page_cgroup_init(pgdat);
4849 for (j = 0; j < MAX_NR_ZONES; j++) {
4850 struct zone *zone = pgdat->node_zones + j;
4851 unsigned long size, realsize, freesize, memmap_pages;
4853 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4854 node_end_pfn, zones_size);
4855 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4861 * Adjust freesize so that it accounts for how much memory
4862 * is used by this zone for memmap. This affects the watermark
4863 * and per-cpu initialisations
4865 memmap_pages = calc_memmap_size(size, realsize);
4866 if (freesize >= memmap_pages) {
4867 freesize -= memmap_pages;
4870 " %s zone: %lu pages used for memmap\n",
4871 zone_names[j], memmap_pages);
4874 " %s zone: %lu pages exceeds freesize %lu\n",
4875 zone_names[j], memmap_pages, freesize);
4877 /* Account for reserved pages */
4878 if (j == 0 && freesize > dma_reserve) {
4879 freesize -= dma_reserve;
4880 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4881 zone_names[0], dma_reserve);
4884 if (!is_highmem_idx(j))
4885 nr_kernel_pages += freesize;
4886 /* Charge for highmem memmap if there are enough kernel pages */
4887 else if (nr_kernel_pages > memmap_pages * 2)
4888 nr_kernel_pages -= memmap_pages;
4889 nr_all_pages += freesize;
4891 zone->spanned_pages = size;
4892 zone->present_pages = realsize;
4894 * Set an approximate value for lowmem here, it will be adjusted
4895 * when the bootmem allocator frees pages into the buddy system.
4896 * And all highmem pages will be managed by the buddy system.
4898 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4901 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4903 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4905 zone->name = zone_names[j];
4906 spin_lock_init(&zone->lock);
4907 spin_lock_init(&zone->lru_lock);
4908 zone_seqlock_init(zone);
4909 zone->zone_pgdat = pgdat;
4910 zone_pcp_init(zone);
4912 /* For bootup, initialized properly in watermark setup */
4913 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4915 lruvec_init(&zone->lruvec);
4919 set_pageblock_order();
4920 setup_usemap(pgdat, zone, zone_start_pfn, size);
4921 ret = init_currently_empty_zone(zone, zone_start_pfn,
4922 size, MEMMAP_EARLY);
4924 memmap_init(size, nid, j, zone_start_pfn);
4925 zone_start_pfn += size;
4929 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4931 /* Skip empty nodes */
4932 if (!pgdat->node_spanned_pages)
4935 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4936 /* ia64 gets its own node_mem_map, before this, without bootmem */
4937 if (!pgdat->node_mem_map) {
4938 unsigned long size, start, end;
4942 * The zone's endpoints aren't required to be MAX_ORDER
4943 * aligned but the node_mem_map endpoints must be in order
4944 * for the buddy allocator to function correctly.
4946 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4947 end = pgdat_end_pfn(pgdat);
4948 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4949 size = (end - start) * sizeof(struct page);
4950 map = alloc_remap(pgdat->node_id, size);
4952 map = memblock_virt_alloc_node_nopanic(size,
4954 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4956 #ifndef CONFIG_NEED_MULTIPLE_NODES
4958 * With no DISCONTIG, the global mem_map is just set as node 0's
4960 if (pgdat == NODE_DATA(0)) {
4961 mem_map = NODE_DATA(0)->node_mem_map;
4962 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4963 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4964 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4965 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4968 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4971 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4972 unsigned long node_start_pfn, unsigned long *zholes_size)
4974 pg_data_t *pgdat = NODE_DATA(nid);
4975 unsigned long start_pfn = 0;
4976 unsigned long end_pfn = 0;
4978 /* pg_data_t should be reset to zero when it's allocated */
4979 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4981 pgdat->node_id = nid;
4982 pgdat->node_start_pfn = node_start_pfn;
4983 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4984 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4985 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid,
4986 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1);
4988 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4989 zones_size, zholes_size);
4991 alloc_node_mem_map(pgdat);
4992 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4993 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4994 nid, (unsigned long)pgdat,
4995 (unsigned long)pgdat->node_mem_map);
4998 free_area_init_core(pgdat, start_pfn, end_pfn,
4999 zones_size, zholes_size);
5002 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5004 #if MAX_NUMNODES > 1
5006 * Figure out the number of possible node ids.
5008 void __init setup_nr_node_ids(void)
5011 unsigned int highest = 0;
5013 for_each_node_mask(node, node_possible_map)
5015 nr_node_ids = highest + 1;
5020 * node_map_pfn_alignment - determine the maximum internode alignment
5022 * This function should be called after node map is populated and sorted.
5023 * It calculates the maximum power of two alignment which can distinguish
5026 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5027 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5028 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5029 * shifted, 1GiB is enough and this function will indicate so.
5031 * This is used to test whether pfn -> nid mapping of the chosen memory
5032 * model has fine enough granularity to avoid incorrect mapping for the
5033 * populated node map.
5035 * Returns the determined alignment in pfn's. 0 if there is no alignment
5036 * requirement (single node).
5038 unsigned long __init node_map_pfn_alignment(void)
5040 unsigned long accl_mask = 0, last_end = 0;
5041 unsigned long start, end, mask;
5045 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5046 if (!start || last_nid < 0 || last_nid == nid) {
5053 * Start with a mask granular enough to pin-point to the
5054 * start pfn and tick off bits one-by-one until it becomes
5055 * too coarse to separate the current node from the last.
5057 mask = ~((1 << __ffs(start)) - 1);
5058 while (mask && last_end <= (start & (mask << 1)))
5061 /* accumulate all internode masks */
5065 /* convert mask to number of pages */
5066 return ~accl_mask + 1;
5069 /* Find the lowest pfn for a node */
5070 static unsigned long __init find_min_pfn_for_node(int nid)
5072 unsigned long min_pfn = ULONG_MAX;
5073 unsigned long start_pfn;
5076 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5077 min_pfn = min(min_pfn, start_pfn);
5079 if (min_pfn == ULONG_MAX) {
5081 "Could not find start_pfn for node %d\n", nid);
5089 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5091 * It returns the minimum PFN based on information provided via
5092 * memblock_set_node().
5094 unsigned long __init find_min_pfn_with_active_regions(void)
5096 return find_min_pfn_for_node(MAX_NUMNODES);
5100 * early_calculate_totalpages()
5101 * Sum pages in active regions for movable zone.
5102 * Populate N_MEMORY for calculating usable_nodes.
5104 static unsigned long __init early_calculate_totalpages(void)
5106 unsigned long totalpages = 0;
5107 unsigned long start_pfn, end_pfn;
5110 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5111 unsigned long pages = end_pfn - start_pfn;
5113 totalpages += pages;
5115 node_set_state(nid, N_MEMORY);
5121 * Find the PFN the Movable zone begins in each node. Kernel memory
5122 * is spread evenly between nodes as long as the nodes have enough
5123 * memory. When they don't, some nodes will have more kernelcore than
5126 static void __init find_zone_movable_pfns_for_nodes(void)
5129 unsigned long usable_startpfn;
5130 unsigned long kernelcore_node, kernelcore_remaining;
5131 /* save the state before borrow the nodemask */
5132 nodemask_t saved_node_state = node_states[N_MEMORY];
5133 unsigned long totalpages = early_calculate_totalpages();
5134 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5135 struct memblock_region *r;
5137 /* Need to find movable_zone earlier when movable_node is specified. */
5138 find_usable_zone_for_movable();
5141 * If movable_node is specified, ignore kernelcore and movablecore
5144 if (movable_node_is_enabled()) {
5145 for_each_memblock(memory, r) {
5146 if (!memblock_is_hotpluggable(r))
5151 usable_startpfn = PFN_DOWN(r->base);
5152 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5153 min(usable_startpfn, zone_movable_pfn[nid]) :
5161 * If movablecore=nn[KMG] was specified, calculate what size of
5162 * kernelcore that corresponds so that memory usable for
5163 * any allocation type is evenly spread. If both kernelcore
5164 * and movablecore are specified, then the value of kernelcore
5165 * will be used for required_kernelcore if it's greater than
5166 * what movablecore would have allowed.
5168 if (required_movablecore) {
5169 unsigned long corepages;
5172 * Round-up so that ZONE_MOVABLE is at least as large as what
5173 * was requested by the user
5175 required_movablecore =
5176 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5177 corepages = totalpages - required_movablecore;
5179 required_kernelcore = max(required_kernelcore, corepages);
5182 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5183 if (!required_kernelcore)
5186 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5187 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5190 /* Spread kernelcore memory as evenly as possible throughout nodes */
5191 kernelcore_node = required_kernelcore / usable_nodes;
5192 for_each_node_state(nid, N_MEMORY) {
5193 unsigned long start_pfn, end_pfn;
5196 * Recalculate kernelcore_node if the division per node
5197 * now exceeds what is necessary to satisfy the requested
5198 * amount of memory for the kernel
5200 if (required_kernelcore < kernelcore_node)
5201 kernelcore_node = required_kernelcore / usable_nodes;
5204 * As the map is walked, we track how much memory is usable
5205 * by the kernel using kernelcore_remaining. When it is
5206 * 0, the rest of the node is usable by ZONE_MOVABLE
5208 kernelcore_remaining = kernelcore_node;
5210 /* Go through each range of PFNs within this node */
5211 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5212 unsigned long size_pages;
5214 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5215 if (start_pfn >= end_pfn)
5218 /* Account for what is only usable for kernelcore */
5219 if (start_pfn < usable_startpfn) {
5220 unsigned long kernel_pages;
5221 kernel_pages = min(end_pfn, usable_startpfn)
5224 kernelcore_remaining -= min(kernel_pages,
5225 kernelcore_remaining);
5226 required_kernelcore -= min(kernel_pages,
5227 required_kernelcore);
5229 /* Continue if range is now fully accounted */
5230 if (end_pfn <= usable_startpfn) {
5233 * Push zone_movable_pfn to the end so
5234 * that if we have to rebalance
5235 * kernelcore across nodes, we will
5236 * not double account here
5238 zone_movable_pfn[nid] = end_pfn;
5241 start_pfn = usable_startpfn;
5245 * The usable PFN range for ZONE_MOVABLE is from
5246 * start_pfn->end_pfn. Calculate size_pages as the
5247 * number of pages used as kernelcore
5249 size_pages = end_pfn - start_pfn;
5250 if (size_pages > kernelcore_remaining)
5251 size_pages = kernelcore_remaining;
5252 zone_movable_pfn[nid] = start_pfn + size_pages;
5255 * Some kernelcore has been met, update counts and
5256 * break if the kernelcore for this node has been
5259 required_kernelcore -= min(required_kernelcore,
5261 kernelcore_remaining -= size_pages;
5262 if (!kernelcore_remaining)
5268 * If there is still required_kernelcore, we do another pass with one
5269 * less node in the count. This will push zone_movable_pfn[nid] further
5270 * along on the nodes that still have memory until kernelcore is
5274 if (usable_nodes && required_kernelcore > usable_nodes)
5278 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5279 for (nid = 0; nid < MAX_NUMNODES; nid++)
5280 zone_movable_pfn[nid] =
5281 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5284 /* restore the node_state */
5285 node_states[N_MEMORY] = saved_node_state;
5288 /* Any regular or high memory on that node ? */
5289 static void check_for_memory(pg_data_t *pgdat, int nid)
5291 enum zone_type zone_type;
5293 if (N_MEMORY == N_NORMAL_MEMORY)
5296 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5297 struct zone *zone = &pgdat->node_zones[zone_type];
5298 if (populated_zone(zone)) {
5299 node_set_state(nid, N_HIGH_MEMORY);
5300 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5301 zone_type <= ZONE_NORMAL)
5302 node_set_state(nid, N_NORMAL_MEMORY);
5309 * free_area_init_nodes - Initialise all pg_data_t and zone data
5310 * @max_zone_pfn: an array of max PFNs for each zone
5312 * This will call free_area_init_node() for each active node in the system.
5313 * Using the page ranges provided by memblock_set_node(), the size of each
5314 * zone in each node and their holes is calculated. If the maximum PFN
5315 * between two adjacent zones match, it is assumed that the zone is empty.
5316 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5317 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5318 * starts where the previous one ended. For example, ZONE_DMA32 starts
5319 * at arch_max_dma_pfn.
5321 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5323 unsigned long start_pfn, end_pfn;
5326 /* Record where the zone boundaries are */
5327 memset(arch_zone_lowest_possible_pfn, 0,
5328 sizeof(arch_zone_lowest_possible_pfn));
5329 memset(arch_zone_highest_possible_pfn, 0,
5330 sizeof(arch_zone_highest_possible_pfn));
5331 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5332 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5333 for (i = 1; i < MAX_NR_ZONES; i++) {
5334 if (i == ZONE_MOVABLE)
5336 arch_zone_lowest_possible_pfn[i] =
5337 arch_zone_highest_possible_pfn[i-1];
5338 arch_zone_highest_possible_pfn[i] =
5339 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5341 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5342 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5344 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5345 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5346 find_zone_movable_pfns_for_nodes();
5348 /* Print out the zone ranges */
5349 printk("Zone ranges:\n");
5350 for (i = 0; i < MAX_NR_ZONES; i++) {
5351 if (i == ZONE_MOVABLE)
5353 printk(KERN_CONT " %-8s ", zone_names[i]);
5354 if (arch_zone_lowest_possible_pfn[i] ==
5355 arch_zone_highest_possible_pfn[i])
5356 printk(KERN_CONT "empty\n");
5358 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5359 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5360 (arch_zone_highest_possible_pfn[i]
5361 << PAGE_SHIFT) - 1);
5364 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5365 printk("Movable zone start for each node\n");
5366 for (i = 0; i < MAX_NUMNODES; i++) {
5367 if (zone_movable_pfn[i])
5368 printk(" Node %d: %#010lx\n", i,
5369 zone_movable_pfn[i] << PAGE_SHIFT);
5372 /* Print out the early node map */
5373 printk("Early memory node ranges\n");
5374 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5375 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5376 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5378 /* Initialise every node */
5379 mminit_verify_pageflags_layout();
5380 setup_nr_node_ids();
5381 for_each_online_node(nid) {
5382 pg_data_t *pgdat = NODE_DATA(nid);
5383 free_area_init_node(nid, NULL,
5384 find_min_pfn_for_node(nid), NULL);
5386 /* Any memory on that node */
5387 if (pgdat->node_present_pages)
5388 node_set_state(nid, N_MEMORY);
5389 check_for_memory(pgdat, nid);
5393 static int __init cmdline_parse_core(char *p, unsigned long *core)
5395 unsigned long long coremem;
5399 coremem = memparse(p, &p);
5400 *core = coremem >> PAGE_SHIFT;
5402 /* Paranoid check that UL is enough for the coremem value */
5403 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5409 * kernelcore=size sets the amount of memory for use for allocations that
5410 * cannot be reclaimed or migrated.
5412 static int __init cmdline_parse_kernelcore(char *p)
5414 return cmdline_parse_core(p, &required_kernelcore);
5418 * movablecore=size sets the amount of memory for use for allocations that
5419 * can be reclaimed or migrated.
5421 static int __init cmdline_parse_movablecore(char *p)
5423 return cmdline_parse_core(p, &required_movablecore);
5426 early_param("kernelcore", cmdline_parse_kernelcore);
5427 early_param("movablecore", cmdline_parse_movablecore);
5429 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5431 void adjust_managed_page_count(struct page *page, long count)
5433 spin_lock(&managed_page_count_lock);
5434 page_zone(page)->managed_pages += count;
5435 totalram_pages += count;
5436 #ifdef CONFIG_HIGHMEM
5437 if (PageHighMem(page))
5438 totalhigh_pages += count;
5440 spin_unlock(&managed_page_count_lock);
5442 EXPORT_SYMBOL(adjust_managed_page_count);
5444 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5447 unsigned long pages = 0;
5449 start = (void *)PAGE_ALIGN((unsigned long)start);
5450 end = (void *)((unsigned long)end & PAGE_MASK);
5451 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5452 if ((unsigned int)poison <= 0xFF)
5453 memset(pos, poison, PAGE_SIZE);
5454 free_reserved_page(virt_to_page(pos));
5458 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5459 s, pages << (PAGE_SHIFT - 10), start, end);
5463 EXPORT_SYMBOL(free_reserved_area);
5465 #ifdef CONFIG_HIGHMEM
5466 void free_highmem_page(struct page *page)
5468 __free_reserved_page(page);
5470 page_zone(page)->managed_pages++;
5476 void __init mem_init_print_info(const char *str)
5478 unsigned long physpages, codesize, datasize, rosize, bss_size;
5479 unsigned long init_code_size, init_data_size;
5481 physpages = get_num_physpages();
5482 codesize = _etext - _stext;
5483 datasize = _edata - _sdata;
5484 rosize = __end_rodata - __start_rodata;
5485 bss_size = __bss_stop - __bss_start;
5486 init_data_size = __init_end - __init_begin;
5487 init_code_size = _einittext - _sinittext;
5490 * Detect special cases and adjust section sizes accordingly:
5491 * 1) .init.* may be embedded into .data sections
5492 * 2) .init.text.* may be out of [__init_begin, __init_end],
5493 * please refer to arch/tile/kernel/vmlinux.lds.S.
5494 * 3) .rodata.* may be embedded into .text or .data sections.
5496 #define adj_init_size(start, end, size, pos, adj) \
5498 if (start <= pos && pos < end && size > adj) \
5502 adj_init_size(__init_begin, __init_end, init_data_size,
5503 _sinittext, init_code_size);
5504 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5505 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5506 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5507 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5509 #undef adj_init_size
5511 printk("Memory: %luK/%luK available "
5512 "(%luK kernel code, %luK rwdata, %luK rodata, "
5513 "%luK init, %luK bss, %luK reserved"
5514 #ifdef CONFIG_HIGHMEM
5518 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5519 codesize >> 10, datasize >> 10, rosize >> 10,
5520 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5521 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5522 #ifdef CONFIG_HIGHMEM
5523 totalhigh_pages << (PAGE_SHIFT-10),
5525 str ? ", " : "", str ? str : "");
5529 * set_dma_reserve - set the specified number of pages reserved in the first zone
5530 * @new_dma_reserve: The number of pages to mark reserved
5532 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5533 * In the DMA zone, a significant percentage may be consumed by kernel image
5534 * and other unfreeable allocations which can skew the watermarks badly. This
5535 * function may optionally be used to account for unfreeable pages in the
5536 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5537 * smaller per-cpu batchsize.
5539 void __init set_dma_reserve(unsigned long new_dma_reserve)
5541 dma_reserve = new_dma_reserve;
5544 void __init free_area_init(unsigned long *zones_size)
5546 free_area_init_node(0, zones_size,
5547 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5550 static int page_alloc_cpu_notify(struct notifier_block *self,
5551 unsigned long action, void *hcpu)
5553 int cpu = (unsigned long)hcpu;
5555 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5556 lru_add_drain_cpu(cpu);
5560 * Spill the event counters of the dead processor
5561 * into the current processors event counters.
5562 * This artificially elevates the count of the current
5565 vm_events_fold_cpu(cpu);
5568 * Zero the differential counters of the dead processor
5569 * so that the vm statistics are consistent.
5571 * This is only okay since the processor is dead and cannot
5572 * race with what we are doing.
5574 cpu_vm_stats_fold(cpu);
5579 void __init page_alloc_init(void)
5581 hotcpu_notifier(page_alloc_cpu_notify, 0);
5585 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5586 * or min_free_kbytes changes.
5588 static void calculate_totalreserve_pages(void)
5590 struct pglist_data *pgdat;
5591 unsigned long reserve_pages = 0;
5592 enum zone_type i, j;
5594 for_each_online_pgdat(pgdat) {
5595 for (i = 0; i < MAX_NR_ZONES; i++) {
5596 struct zone *zone = pgdat->node_zones + i;
5599 /* Find valid and maximum lowmem_reserve in the zone */
5600 for (j = i; j < MAX_NR_ZONES; j++) {
5601 if (zone->lowmem_reserve[j] > max)
5602 max = zone->lowmem_reserve[j];
5605 /* we treat the high watermark as reserved pages. */
5606 max += high_wmark_pages(zone);
5608 if (max > zone->managed_pages)
5609 max = zone->managed_pages;
5610 reserve_pages += max;
5612 * Lowmem reserves are not available to
5613 * GFP_HIGHUSER page cache allocations and
5614 * kswapd tries to balance zones to their high
5615 * watermark. As a result, neither should be
5616 * regarded as dirtyable memory, to prevent a
5617 * situation where reclaim has to clean pages
5618 * in order to balance the zones.
5620 zone->dirty_balance_reserve = max;
5623 dirty_balance_reserve = reserve_pages;
5624 totalreserve_pages = reserve_pages;
5628 * setup_per_zone_lowmem_reserve - called whenever
5629 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5630 * has a correct pages reserved value, so an adequate number of
5631 * pages are left in the zone after a successful __alloc_pages().
5633 static void setup_per_zone_lowmem_reserve(void)
5635 struct pglist_data *pgdat;
5636 enum zone_type j, idx;
5638 for_each_online_pgdat(pgdat) {
5639 for (j = 0; j < MAX_NR_ZONES; j++) {
5640 struct zone *zone = pgdat->node_zones + j;
5641 unsigned long managed_pages = zone->managed_pages;
5643 zone->lowmem_reserve[j] = 0;
5647 struct zone *lower_zone;
5651 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5652 sysctl_lowmem_reserve_ratio[idx] = 1;
5654 lower_zone = pgdat->node_zones + idx;
5655 lower_zone->lowmem_reserve[j] = managed_pages /
5656 sysctl_lowmem_reserve_ratio[idx];
5657 managed_pages += lower_zone->managed_pages;
5662 /* update totalreserve_pages */
5663 calculate_totalreserve_pages();
5666 static void __setup_per_zone_wmarks(void)
5668 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5669 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
5670 unsigned long lowmem_pages = 0;
5672 unsigned long flags;
5674 /* Calculate total number of !ZONE_HIGHMEM pages */
5675 for_each_zone(zone) {
5676 if (!is_highmem(zone))
5677 lowmem_pages += zone->managed_pages;
5680 for_each_zone(zone) {
5683 spin_lock_irqsave(&zone->lock, flags);
5684 min = (u64)pages_min * zone->managed_pages;
5685 do_div(min, lowmem_pages);
5686 low = (u64)pages_low * zone->managed_pages;
5687 do_div(low, vm_total_pages);
5689 if (is_highmem(zone)) {
5691 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5692 * need highmem pages, so cap pages_min to a small
5695 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5696 * deltas controls asynch page reclaim, and so should
5697 * not be capped for highmem.
5699 unsigned long min_pages;
5701 min_pages = zone->managed_pages / 1024;
5702 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5703 zone->watermark[WMARK_MIN] = min_pages;
5706 * If it's a lowmem zone, reserve a number of pages
5707 * proportionate to the zone's size.
5709 zone->watermark[WMARK_MIN] = min;
5712 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
5714 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
5717 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5718 high_wmark_pages(zone) - low_wmark_pages(zone) -
5719 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5721 setup_zone_migrate_reserve(zone);
5722 spin_unlock_irqrestore(&zone->lock, flags);
5725 /* update totalreserve_pages */
5726 calculate_totalreserve_pages();
5730 * setup_per_zone_wmarks - called when min_free_kbytes changes
5731 * or when memory is hot-{added|removed}
5733 * Ensures that the watermark[min,low,high] values for each zone are set
5734 * correctly with respect to min_free_kbytes.
5736 void setup_per_zone_wmarks(void)
5738 mutex_lock(&zonelists_mutex);
5739 __setup_per_zone_wmarks();
5740 mutex_unlock(&zonelists_mutex);
5744 * The inactive anon list should be small enough that the VM never has to
5745 * do too much work, but large enough that each inactive page has a chance
5746 * to be referenced again before it is swapped out.
5748 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5749 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5750 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5751 * the anonymous pages are kept on the inactive list.
5754 * memory ratio inactive anon
5755 * -------------------------------------
5764 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5766 unsigned int gb, ratio;
5768 /* Zone size in gigabytes */
5769 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5771 ratio = int_sqrt(10 * gb);
5775 zone->inactive_ratio = ratio;
5778 static void __meminit setup_per_zone_inactive_ratio(void)
5783 calculate_zone_inactive_ratio(zone);
5787 * Initialise min_free_kbytes.
5789 * For small machines we want it small (128k min). For large machines
5790 * we want it large (64MB max). But it is not linear, because network
5791 * bandwidth does not increase linearly with machine size. We use
5793 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5794 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5810 int __meminit init_per_zone_wmark_min(void)
5812 unsigned long lowmem_kbytes;
5813 int new_min_free_kbytes;
5815 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5816 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5818 if (new_min_free_kbytes > user_min_free_kbytes) {
5819 min_free_kbytes = new_min_free_kbytes;
5820 if (min_free_kbytes < 128)
5821 min_free_kbytes = 128;
5822 if (min_free_kbytes > 65536)
5823 min_free_kbytes = 65536;
5825 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5826 new_min_free_kbytes, user_min_free_kbytes);
5828 setup_per_zone_wmarks();
5829 refresh_zone_stat_thresholds();
5830 setup_per_zone_lowmem_reserve();
5831 setup_per_zone_inactive_ratio();
5834 module_init(init_per_zone_wmark_min)
5837 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5838 * that we can call two helper functions whenever min_free_kbytes
5839 * or extra_free_kbytes changes.
5841 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5842 void __user *buffer, size_t *length, loff_t *ppos)
5846 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5851 user_min_free_kbytes = min_free_kbytes;
5852 setup_per_zone_wmarks();
5858 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5859 void __user *buffer, size_t *length, loff_t *ppos)
5864 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5869 zone->min_unmapped_pages = (zone->managed_pages *
5870 sysctl_min_unmapped_ratio) / 100;
5874 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5875 void __user *buffer, size_t *length, loff_t *ppos)
5880 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5885 zone->min_slab_pages = (zone->managed_pages *
5886 sysctl_min_slab_ratio) / 100;
5892 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5893 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5894 * whenever sysctl_lowmem_reserve_ratio changes.
5896 * The reserve ratio obviously has absolutely no relation with the
5897 * minimum watermarks. The lowmem reserve ratio can only make sense
5898 * if in function of the boot time zone sizes.
5900 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5901 void __user *buffer, size_t *length, loff_t *ppos)
5903 proc_dointvec_minmax(table, write, buffer, length, ppos);
5904 setup_per_zone_lowmem_reserve();
5909 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5910 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5911 * pagelist can have before it gets flushed back to buddy allocator.
5913 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5914 void __user *buffer, size_t *length, loff_t *ppos)
5917 int old_percpu_pagelist_fraction;
5920 mutex_lock(&pcp_batch_high_lock);
5921 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5923 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5924 if (!write || ret < 0)
5927 /* Sanity checking to avoid pcp imbalance */
5928 if (percpu_pagelist_fraction &&
5929 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5930 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5936 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5939 for_each_populated_zone(zone) {
5942 for_each_possible_cpu(cpu)
5943 pageset_set_high_and_batch(zone,
5944 per_cpu_ptr(zone->pageset, cpu));
5947 mutex_unlock(&pcp_batch_high_lock);
5951 int hashdist = HASHDIST_DEFAULT;
5954 static int __init set_hashdist(char *str)
5958 hashdist = simple_strtoul(str, &str, 0);
5961 __setup("hashdist=", set_hashdist);
5965 * allocate a large system hash table from bootmem
5966 * - it is assumed that the hash table must contain an exact power-of-2
5967 * quantity of entries
5968 * - limit is the number of hash buckets, not the total allocation size
5970 void *__init alloc_large_system_hash(const char *tablename,
5971 unsigned long bucketsize,
5972 unsigned long numentries,
5975 unsigned int *_hash_shift,
5976 unsigned int *_hash_mask,
5977 unsigned long low_limit,
5978 unsigned long high_limit)
5980 unsigned long long max = high_limit;
5981 unsigned long log2qty, size;
5984 /* allow the kernel cmdline to have a say */
5986 /* round applicable memory size up to nearest megabyte */
5987 numentries = nr_kernel_pages;
5989 /* It isn't necessary when PAGE_SIZE >= 1MB */
5990 if (PAGE_SHIFT < 20)
5991 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5993 /* limit to 1 bucket per 2^scale bytes of low memory */
5994 if (scale > PAGE_SHIFT)
5995 numentries >>= (scale - PAGE_SHIFT);
5997 numentries <<= (PAGE_SHIFT - scale);
5999 /* Make sure we've got at least a 0-order allocation.. */
6000 if (unlikely(flags & HASH_SMALL)) {
6001 /* Makes no sense without HASH_EARLY */
6002 WARN_ON(!(flags & HASH_EARLY));
6003 if (!(numentries >> *_hash_shift)) {
6004 numentries = 1UL << *_hash_shift;
6005 BUG_ON(!numentries);
6007 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6008 numentries = PAGE_SIZE / bucketsize;
6010 numentries = roundup_pow_of_two(numentries);
6012 /* limit allocation size to 1/16 total memory by default */
6014 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6015 do_div(max, bucketsize);
6017 max = min(max, 0x80000000ULL);
6019 if (numentries < low_limit)
6020 numentries = low_limit;
6021 if (numentries > max)
6024 log2qty = ilog2(numentries);
6027 size = bucketsize << log2qty;
6028 if (flags & HASH_EARLY)
6029 table = memblock_virt_alloc_nopanic(size, 0);
6031 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6034 * If bucketsize is not a power-of-two, we may free
6035 * some pages at the end of hash table which
6036 * alloc_pages_exact() automatically does
6038 if (get_order(size) < MAX_ORDER) {
6039 table = alloc_pages_exact(size, GFP_ATOMIC);
6040 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6043 } while (!table && size > PAGE_SIZE && --log2qty);
6046 panic("Failed to allocate %s hash table\n", tablename);
6048 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6051 ilog2(size) - PAGE_SHIFT,
6055 *_hash_shift = log2qty;
6057 *_hash_mask = (1 << log2qty) - 1;
6062 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6063 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6066 #ifdef CONFIG_SPARSEMEM
6067 return __pfn_to_section(pfn)->pageblock_flags;
6069 return zone->pageblock_flags;
6070 #endif /* CONFIG_SPARSEMEM */
6073 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6075 #ifdef CONFIG_SPARSEMEM
6076 pfn &= (PAGES_PER_SECTION-1);
6077 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6079 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6080 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6081 #endif /* CONFIG_SPARSEMEM */
6085 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6086 * @page: The page within the block of interest
6087 * @pfn: The target page frame number
6088 * @end_bitidx: The last bit of interest to retrieve
6089 * @mask: mask of bits that the caller is interested in
6091 * Return: pageblock_bits flags
6093 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6094 unsigned long end_bitidx,
6098 unsigned long *bitmap;
6099 unsigned long bitidx, word_bitidx;
6102 zone = page_zone(page);
6103 bitmap = get_pageblock_bitmap(zone, pfn);
6104 bitidx = pfn_to_bitidx(zone, pfn);
6105 word_bitidx = bitidx / BITS_PER_LONG;
6106 bitidx &= (BITS_PER_LONG-1);
6108 word = bitmap[word_bitidx];
6109 bitidx += end_bitidx;
6110 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6114 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6115 * @page: The page within the block of interest
6116 * @flags: The flags to set
6117 * @pfn: The target page frame number
6118 * @end_bitidx: The last bit of interest
6119 * @mask: mask of bits that the caller is interested in
6121 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6123 unsigned long end_bitidx,
6127 unsigned long *bitmap;
6128 unsigned long bitidx, word_bitidx;
6129 unsigned long old_word, word;
6131 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6133 zone = page_zone(page);
6134 bitmap = get_pageblock_bitmap(zone, pfn);
6135 bitidx = pfn_to_bitidx(zone, pfn);
6136 word_bitidx = bitidx / BITS_PER_LONG;
6137 bitidx &= (BITS_PER_LONG-1);
6139 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6141 bitidx += end_bitidx;
6142 mask <<= (BITS_PER_LONG - bitidx - 1);
6143 flags <<= (BITS_PER_LONG - bitidx - 1);
6145 word = ACCESS_ONCE(bitmap[word_bitidx]);
6147 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6148 if (word == old_word)
6155 * This function checks whether pageblock includes unmovable pages or not.
6156 * If @count is not zero, it is okay to include less @count unmovable pages
6158 * PageLRU check without isolation or lru_lock could race so that
6159 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6160 * expect this function should be exact.
6162 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6163 bool skip_hwpoisoned_pages)
6165 unsigned long pfn, iter, found;
6169 * For avoiding noise data, lru_add_drain_all() should be called
6170 * If ZONE_MOVABLE, the zone never contains unmovable pages
6172 if (zone_idx(zone) == ZONE_MOVABLE)
6174 mt = get_pageblock_migratetype(page);
6175 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6178 pfn = page_to_pfn(page);
6179 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6180 unsigned long check = pfn + iter;
6182 if (!pfn_valid_within(check))
6185 page = pfn_to_page(check);
6188 * Hugepages are not in LRU lists, but they're movable.
6189 * We need not scan over tail pages bacause we don't
6190 * handle each tail page individually in migration.
6192 if (PageHuge(page)) {
6193 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6198 * We can't use page_count without pin a page
6199 * because another CPU can free compound page.
6200 * This check already skips compound tails of THP
6201 * because their page->_count is zero at all time.
6203 if (!atomic_read(&page->_count)) {
6204 if (PageBuddy(page))
6205 iter += (1 << page_order(page)) - 1;
6210 * The HWPoisoned page may be not in buddy system, and
6211 * page_count() is not 0.
6213 if (skip_hwpoisoned_pages && PageHWPoison(page))
6219 * If there are RECLAIMABLE pages, we need to check it.
6220 * But now, memory offline itself doesn't call shrink_slab()
6221 * and it still to be fixed.
6224 * If the page is not RAM, page_count()should be 0.
6225 * we don't need more check. This is an _used_ not-movable page.
6227 * The problematic thing here is PG_reserved pages. PG_reserved
6228 * is set to both of a memory hole page and a _used_ kernel
6237 bool is_pageblock_removable_nolock(struct page *page)
6243 * We have to be careful here because we are iterating over memory
6244 * sections which are not zone aware so we might end up outside of
6245 * the zone but still within the section.
6246 * We have to take care about the node as well. If the node is offline
6247 * its NODE_DATA will be NULL - see page_zone.
6249 if (!node_online(page_to_nid(page)))
6252 zone = page_zone(page);
6253 pfn = page_to_pfn(page);
6254 if (!zone_spans_pfn(zone, pfn))
6257 return !has_unmovable_pages(zone, page, 0, true);
6262 static unsigned long pfn_max_align_down(unsigned long pfn)
6264 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6265 pageblock_nr_pages) - 1);
6268 static unsigned long pfn_max_align_up(unsigned long pfn)
6270 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6271 pageblock_nr_pages));
6274 /* [start, end) must belong to a single zone. */
6275 static int __alloc_contig_migrate_range(struct compact_control *cc,
6276 unsigned long start, unsigned long end)
6278 /* This function is based on compact_zone() from compaction.c. */
6279 unsigned long nr_reclaimed;
6280 unsigned long pfn = start;
6281 unsigned int tries = 0;
6286 while (pfn < end || !list_empty(&cc->migratepages)) {
6287 if (fatal_signal_pending(current)) {
6292 if (list_empty(&cc->migratepages)) {
6293 cc->nr_migratepages = 0;
6294 pfn = isolate_migratepages_range(cc, pfn, end);
6300 } else if (++tries == 5) {
6301 ret = ret < 0 ? ret : -EBUSY;
6305 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6307 cc->nr_migratepages -= nr_reclaimed;
6309 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6310 NULL, 0, cc->mode, MR_CMA);
6313 putback_movable_pages(&cc->migratepages);
6320 * alloc_contig_range() -- tries to allocate given range of pages
6321 * @start: start PFN to allocate
6322 * @end: one-past-the-last PFN to allocate
6323 * @migratetype: migratetype of the underlaying pageblocks (either
6324 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6325 * in range must have the same migratetype and it must
6326 * be either of the two.
6328 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6329 * aligned, however it's the caller's responsibility to guarantee that
6330 * we are the only thread that changes migrate type of pageblocks the
6333 * The PFN range must belong to a single zone.
6335 * Returns zero on success or negative error code. On success all
6336 * pages which PFN is in [start, end) are allocated for the caller and
6337 * need to be freed with free_contig_range().
6339 int alloc_contig_range(unsigned long start, unsigned long end,
6340 unsigned migratetype)
6342 unsigned long outer_start, outer_end;
6345 struct compact_control cc = {
6346 .nr_migratepages = 0,
6348 .zone = page_zone(pfn_to_page(start)),
6349 .mode = MIGRATE_SYNC,
6350 .ignore_skip_hint = true,
6352 INIT_LIST_HEAD(&cc.migratepages);
6355 * What we do here is we mark all pageblocks in range as
6356 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6357 * have different sizes, and due to the way page allocator
6358 * work, we align the range to biggest of the two pages so
6359 * that page allocator won't try to merge buddies from
6360 * different pageblocks and change MIGRATE_ISOLATE to some
6361 * other migration type.
6363 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6364 * migrate the pages from an unaligned range (ie. pages that
6365 * we are interested in). This will put all the pages in
6366 * range back to page allocator as MIGRATE_ISOLATE.
6368 * When this is done, we take the pages in range from page
6369 * allocator removing them from the buddy system. This way
6370 * page allocator will never consider using them.
6372 * This lets us mark the pageblocks back as
6373 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6374 * aligned range but not in the unaligned, original range are
6375 * put back to page allocator so that buddy can use them.
6378 ret = start_isolate_page_range(pfn_max_align_down(start),
6379 pfn_max_align_up(end), migratetype,
6384 ret = __alloc_contig_migrate_range(&cc, start, end);
6389 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6390 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6391 * more, all pages in [start, end) are free in page allocator.
6392 * What we are going to do is to allocate all pages from
6393 * [start, end) (that is remove them from page allocator).
6395 * The only problem is that pages at the beginning and at the
6396 * end of interesting range may be not aligned with pages that
6397 * page allocator holds, ie. they can be part of higher order
6398 * pages. Because of this, we reserve the bigger range and
6399 * once this is done free the pages we are not interested in.
6401 * We don't have to hold zone->lock here because the pages are
6402 * isolated thus they won't get removed from buddy.
6405 lru_add_drain_all();
6409 outer_start = start;
6410 while (!PageBuddy(pfn_to_page(outer_start))) {
6411 if (++order >= MAX_ORDER) {
6415 outer_start &= ~0UL << order;
6418 /* Make sure the range is really isolated. */
6419 if (test_pages_isolated(outer_start, end, false)) {
6420 pr_info("%s: [%lx, %lx) PFNs busy\n",
6421 __func__, outer_start, end);
6426 /* Grab isolated pages from freelists. */
6427 outer_end = isolate_freepages_range(&cc, outer_start, end);
6433 /* Free head and tail (if any) */
6434 if (start != outer_start)
6435 free_contig_range(outer_start, start - outer_start);
6436 if (end != outer_end)
6437 free_contig_range(end, outer_end - end);
6440 undo_isolate_page_range(pfn_max_align_down(start),
6441 pfn_max_align_up(end), migratetype);
6445 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6447 unsigned int count = 0;
6449 for (; nr_pages--; pfn++) {
6450 struct page *page = pfn_to_page(pfn);
6452 count += page_count(page) != 1;
6455 WARN(count != 0, "%d pages are still in use!\n", count);
6459 #ifdef CONFIG_MEMORY_HOTPLUG
6461 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6462 * page high values need to be recalulated.
6464 void __meminit zone_pcp_update(struct zone *zone)
6467 mutex_lock(&pcp_batch_high_lock);
6468 for_each_possible_cpu(cpu)
6469 pageset_set_high_and_batch(zone,
6470 per_cpu_ptr(zone->pageset, cpu));
6471 mutex_unlock(&pcp_batch_high_lock);
6475 void zone_pcp_reset(struct zone *zone)
6477 unsigned long flags;
6479 struct per_cpu_pageset *pset;
6481 /* avoid races with drain_pages() */
6482 local_irq_save(flags);
6483 if (zone->pageset != &boot_pageset) {
6484 for_each_online_cpu(cpu) {
6485 pset = per_cpu_ptr(zone->pageset, cpu);
6486 drain_zonestat(zone, pset);
6488 free_percpu(zone->pageset);
6489 zone->pageset = &boot_pageset;
6491 local_irq_restore(flags);
6494 #ifdef CONFIG_MEMORY_HOTREMOVE
6496 * All pages in the range must be isolated before calling this.
6499 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6503 unsigned int order, i;
6505 unsigned long flags;
6506 /* find the first valid pfn */
6507 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6512 zone = page_zone(pfn_to_page(pfn));
6513 spin_lock_irqsave(&zone->lock, flags);
6515 while (pfn < end_pfn) {
6516 if (!pfn_valid(pfn)) {
6520 page = pfn_to_page(pfn);
6522 * The HWPoisoned page may be not in buddy system, and
6523 * page_count() is not 0.
6525 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6527 SetPageReserved(page);
6531 BUG_ON(page_count(page));
6532 BUG_ON(!PageBuddy(page));
6533 order = page_order(page);
6534 #ifdef CONFIG_DEBUG_VM
6535 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6536 pfn, 1 << order, end_pfn);
6538 list_del(&page->lru);
6539 rmv_page_order(page);
6540 zone->free_area[order].nr_free--;
6541 for (i = 0; i < (1 << order); i++)
6542 SetPageReserved((page+i));
6543 pfn += (1 << order);
6545 spin_unlock_irqrestore(&zone->lock, flags);
6549 #ifdef CONFIG_MEMORY_FAILURE
6550 bool is_free_buddy_page(struct page *page)
6552 struct zone *zone = page_zone(page);
6553 unsigned long pfn = page_to_pfn(page);
6554 unsigned long flags;
6557 spin_lock_irqsave(&zone->lock, flags);
6558 for (order = 0; order < MAX_ORDER; order++) {
6559 struct page *page_head = page - (pfn & ((1 << order) - 1));
6561 if (PageBuddy(page_head) && page_order(page_head) >= order)
6564 spin_unlock_irqrestore(&zone->lock, flags);
6566 return order < MAX_ORDER;