2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_pages __read_mostly;
115 unsigned long totalcma_pages __read_mostly;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly;
124 int percpu_pagelist_fraction;
125 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex));
142 if (saved_gfp_mask) {
143 gfp_allowed_mask = saved_gfp_mask;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex));
151 WARN_ON(saved_gfp_mask);
152 saved_gfp_mask = gfp_allowed_mask;
153 gfp_allowed_mask &= ~GFP_IOFS;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly;
168 static void __free_pages_ok(struct page *page, unsigned int order);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages);
196 static char * const zone_names[MAX_NR_ZONES] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes = 1024;
211 int user_min_free_kbytes = -1;
213 static unsigned long __meminitdata nr_kernel_pages;
214 static unsigned long __meminitdata nr_all_pages;
215 static unsigned long __meminitdata dma_reserve;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __initdata required_kernelcore;
221 static unsigned long __initdata required_movablecore;
222 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly = MAX_NUMNODES;
231 int nr_online_nodes __read_mostly = 1;
232 EXPORT_SYMBOL(nr_node_ids);
233 EXPORT_SYMBOL(nr_online_nodes);
236 int page_group_by_mobility_disabled __read_mostly;
238 void set_pageblock_migratetype(struct page *page, int migratetype)
240 if (unlikely(page_group_by_mobility_disabled &&
241 migratetype < MIGRATE_PCPTYPES))
242 migratetype = MIGRATE_UNMOVABLE;
244 set_pageblock_flags_group(page, (unsigned long)migratetype,
245 PB_migrate, PB_migrate_end);
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn, zone_to_nid(zone), zone->name,
267 start_pfn, start_pfn + sp);
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page, const char *reason,
301 unsigned long bad_flags)
303 static unsigned long resume;
304 static unsigned long nr_shown;
305 static unsigned long nr_unshown;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page)) {
309 page_mapcount_reset(page); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown == 60) {
318 if (time_before(jiffies, resume)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume = jiffies + 60 * HZ;
333 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
334 current->comm, page_to_pfn(page));
335 dump_page_badflags(page, reason, bad_flags);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page *page)
362 __free_pages_ok(page, compound_order(page));
365 void prep_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
370 set_compound_page_dtor(page, free_compound_page);
371 set_compound_order(page, order);
373 for (i = 1; i < nr_pages; i++) {
374 struct page *p = page + i;
375 set_page_count(p, 0);
376 p->first_page = page;
377 /* Make sure p->first_page is always valid for PageTail() */
383 static inline void prep_zero_page(struct page *page, unsigned int order,
389 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
390 * and __GFP_HIGHMEM from hard or soft interrupt context.
392 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
393 for (i = 0; i < (1 << order); i++)
394 clear_highpage(page + i);
397 #ifdef CONFIG_DEBUG_PAGEALLOC
398 unsigned int _debug_guardpage_minorder;
399 bool _debug_pagealloc_enabled __read_mostly;
400 bool _debug_guardpage_enabled __read_mostly;
402 static int __init early_debug_pagealloc(char *buf)
407 if (strcmp(buf, "on") == 0)
408 _debug_pagealloc_enabled = true;
412 early_param("debug_pagealloc", early_debug_pagealloc);
414 static bool need_debug_guardpage(void)
416 /* If we don't use debug_pagealloc, we don't need guard page */
417 if (!debug_pagealloc_enabled())
423 static void init_debug_guardpage(void)
425 if (!debug_pagealloc_enabled())
428 _debug_guardpage_enabled = true;
431 struct page_ext_operations debug_guardpage_ops = {
432 .need = need_debug_guardpage,
433 .init = init_debug_guardpage,
436 static int __init debug_guardpage_minorder_setup(char *buf)
440 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
441 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
444 _debug_guardpage_minorder = res;
445 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
448 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
450 static inline void set_page_guard(struct zone *zone, struct page *page,
451 unsigned int order, int migratetype)
453 struct page_ext *page_ext;
455 if (!debug_guardpage_enabled())
458 page_ext = lookup_page_ext(page);
459 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
461 INIT_LIST_HEAD(&page->lru);
462 set_page_private(page, order);
463 /* Guard pages are not available for any usage */
464 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
467 static inline void clear_page_guard(struct zone *zone, struct page *page,
468 unsigned int order, int migratetype)
470 struct page_ext *page_ext;
472 if (!debug_guardpage_enabled())
475 page_ext = lookup_page_ext(page);
476 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
478 set_page_private(page, 0);
479 if (!is_migrate_isolate(migratetype))
480 __mod_zone_freepage_state(zone, (1 << order), migratetype);
483 struct page_ext_operations debug_guardpage_ops = { NULL, };
484 static inline void set_page_guard(struct zone *zone, struct page *page,
485 unsigned int order, int migratetype) {}
486 static inline void clear_page_guard(struct zone *zone, struct page *page,
487 unsigned int order, int migratetype) {}
490 static inline void set_page_order(struct page *page, unsigned int order)
492 set_page_private(page, order);
493 __SetPageBuddy(page);
496 static inline void rmv_page_order(struct page *page)
498 __ClearPageBuddy(page);
499 set_page_private(page, 0);
503 * This function checks whether a page is free && is the buddy
504 * we can do coalesce a page and its buddy if
505 * (a) the buddy is not in a hole &&
506 * (b) the buddy is in the buddy system &&
507 * (c) a page and its buddy have the same order &&
508 * (d) a page and its buddy are in the same zone.
510 * For recording whether a page is in the buddy system, we set ->_mapcount
511 * PAGE_BUDDY_MAPCOUNT_VALUE.
512 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
513 * serialized by zone->lock.
515 * For recording page's order, we use page_private(page).
517 static inline int page_is_buddy(struct page *page, struct page *buddy,
520 if (!pfn_valid_within(page_to_pfn(buddy)))
523 if (page_is_guard(buddy) && page_order(buddy) == order) {
524 if (page_zone_id(page) != page_zone_id(buddy))
527 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
532 if (PageBuddy(buddy) && page_order(buddy) == order) {
534 * zone check is done late to avoid uselessly
535 * calculating zone/node ids for pages that could
538 if (page_zone_id(page) != page_zone_id(buddy))
541 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
549 * Freeing function for a buddy system allocator.
551 * The concept of a buddy system is to maintain direct-mapped table
552 * (containing bit values) for memory blocks of various "orders".
553 * The bottom level table contains the map for the smallest allocatable
554 * units of memory (here, pages), and each level above it describes
555 * pairs of units from the levels below, hence, "buddies".
556 * At a high level, all that happens here is marking the table entry
557 * at the bottom level available, and propagating the changes upward
558 * as necessary, plus some accounting needed to play nicely with other
559 * parts of the VM system.
560 * At each level, we keep a list of pages, which are heads of continuous
561 * free pages of length of (1 << order) and marked with _mapcount
562 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
564 * So when we are allocating or freeing one, we can derive the state of the
565 * other. That is, if we allocate a small block, and both were
566 * free, the remainder of the region must be split into blocks.
567 * If a block is freed, and its buddy is also free, then this
568 * triggers coalescing into a block of larger size.
573 static inline void __free_one_page(struct page *page,
575 struct zone *zone, unsigned int order,
578 unsigned long page_idx;
579 unsigned long combined_idx;
580 unsigned long uninitialized_var(buddy_idx);
582 int max_order = MAX_ORDER;
584 VM_BUG_ON(!zone_is_initialized(zone));
585 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
587 VM_BUG_ON(migratetype == -1);
588 if (is_migrate_isolate(migratetype)) {
590 * We restrict max order of merging to prevent merge
591 * between freepages on isolate pageblock and normal
592 * pageblock. Without this, pageblock isolation
593 * could cause incorrect freepage accounting.
595 max_order = min(MAX_ORDER, pageblock_order + 1);
597 __mod_zone_freepage_state(zone, 1 << order, migratetype);
600 page_idx = pfn & ((1 << max_order) - 1);
602 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
603 VM_BUG_ON_PAGE(bad_range(zone, page), page);
605 while (order < max_order - 1) {
606 buddy_idx = __find_buddy_index(page_idx, order);
607 buddy = page + (buddy_idx - page_idx);
608 if (!page_is_buddy(page, buddy, order))
611 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
612 * merge with it and move up one order.
614 if (page_is_guard(buddy)) {
615 clear_page_guard(zone, buddy, order, migratetype);
617 list_del(&buddy->lru);
618 zone->free_area[order].nr_free--;
619 rmv_page_order(buddy);
621 combined_idx = buddy_idx & page_idx;
622 page = page + (combined_idx - page_idx);
623 page_idx = combined_idx;
626 set_page_order(page, order);
629 * If this is not the largest possible page, check if the buddy
630 * of the next-highest order is free. If it is, it's possible
631 * that pages are being freed that will coalesce soon. In case,
632 * that is happening, add the free page to the tail of the list
633 * so it's less likely to be used soon and more likely to be merged
634 * as a higher order page
636 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
637 struct page *higher_page, *higher_buddy;
638 combined_idx = buddy_idx & page_idx;
639 higher_page = page + (combined_idx - page_idx);
640 buddy_idx = __find_buddy_index(combined_idx, order + 1);
641 higher_buddy = higher_page + (buddy_idx - combined_idx);
642 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
643 list_add_tail(&page->lru,
644 &zone->free_area[order].free_list[migratetype]);
649 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
651 zone->free_area[order].nr_free++;
654 static inline int free_pages_check(struct page *page)
656 const char *bad_reason = NULL;
657 unsigned long bad_flags = 0;
659 if (unlikely(page_mapcount(page)))
660 bad_reason = "nonzero mapcount";
661 if (unlikely(page->mapping != NULL))
662 bad_reason = "non-NULL mapping";
663 if (unlikely(atomic_read(&page->_count) != 0))
664 bad_reason = "nonzero _count";
665 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
666 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
667 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
670 if (unlikely(page->mem_cgroup))
671 bad_reason = "page still charged to cgroup";
673 if (unlikely(bad_reason)) {
674 bad_page(page, bad_reason, bad_flags);
677 page_cpupid_reset_last(page);
678 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
679 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
684 * Frees a number of pages from the PCP lists
685 * Assumes all pages on list are in same zone, and of same order.
686 * count is the number of pages to free.
688 * If the zone was previously in an "all pages pinned" state then look to
689 * see if this freeing clears that state.
691 * And clear the zone's pages_scanned counter, to hold off the "all pages are
692 * pinned" detection logic.
694 static void free_pcppages_bulk(struct zone *zone, int count,
695 struct per_cpu_pages *pcp)
700 unsigned long nr_scanned;
702 spin_lock(&zone->lock);
703 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
705 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
709 struct list_head *list;
712 * Remove pages from lists in a round-robin fashion. A
713 * batch_free count is maintained that is incremented when an
714 * empty list is encountered. This is so more pages are freed
715 * off fuller lists instead of spinning excessively around empty
720 if (++migratetype == MIGRATE_PCPTYPES)
722 list = &pcp->lists[migratetype];
723 } while (list_empty(list));
725 /* This is the only non-empty list. Free them all. */
726 if (batch_free == MIGRATE_PCPTYPES)
727 batch_free = to_free;
730 int mt; /* migratetype of the to-be-freed page */
732 page = list_entry(list->prev, struct page, lru);
733 /* must delete as __free_one_page list manipulates */
734 list_del(&page->lru);
735 mt = get_freepage_migratetype(page);
736 if (unlikely(has_isolate_pageblock(zone)))
737 mt = get_pageblock_migratetype(page);
739 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
740 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
741 trace_mm_page_pcpu_drain(page, 0, mt);
742 } while (--to_free && --batch_free && !list_empty(list));
744 spin_unlock(&zone->lock);
747 static void free_one_page(struct zone *zone,
748 struct page *page, unsigned long pfn,
752 unsigned long nr_scanned;
753 spin_lock(&zone->lock);
754 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
756 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
758 if (unlikely(has_isolate_pageblock(zone) ||
759 is_migrate_isolate(migratetype))) {
760 migratetype = get_pfnblock_migratetype(page, pfn);
762 __free_one_page(page, pfn, zone, order, migratetype);
763 spin_unlock(&zone->lock);
766 static int free_tail_pages_check(struct page *head_page, struct page *page)
768 if (!IS_ENABLED(CONFIG_DEBUG_VM))
770 if (unlikely(!PageTail(page))) {
771 bad_page(page, "PageTail not set", 0);
774 if (unlikely(page->first_page != head_page)) {
775 bad_page(page, "first_page not consistent", 0);
781 static bool free_pages_prepare(struct page *page, unsigned int order)
783 bool compound = PageCompound(page);
786 VM_BUG_ON_PAGE(PageTail(page), page);
787 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
789 trace_mm_page_free(page, order);
790 kmemcheck_free_shadow(page, order);
791 kasan_free_pages(page, order);
794 page->mapping = NULL;
795 bad += free_pages_check(page);
796 for (i = 1; i < (1 << order); i++) {
798 bad += free_tail_pages_check(page, page + i);
799 bad += free_pages_check(page + i);
804 reset_page_owner(page, order);
806 if (!PageHighMem(page)) {
807 debug_check_no_locks_freed(page_address(page),
809 debug_check_no_obj_freed(page_address(page),
812 arch_free_page(page, order);
813 kernel_map_pages(page, 1 << order, 0);
818 static void __free_pages_ok(struct page *page, unsigned int order)
822 unsigned long pfn = page_to_pfn(page);
824 if (!free_pages_prepare(page, order))
827 migratetype = get_pfnblock_migratetype(page, pfn);
828 local_irq_save(flags);
829 __count_vm_events(PGFREE, 1 << order);
830 set_freepage_migratetype(page, migratetype);
831 free_one_page(page_zone(page), page, pfn, order, migratetype);
832 local_irq_restore(flags);
835 void __init __free_pages_bootmem(struct page *page, unsigned int order)
837 unsigned int nr_pages = 1 << order;
838 struct page *p = page;
842 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
844 __ClearPageReserved(p);
845 set_page_count(p, 0);
847 __ClearPageReserved(p);
848 set_page_count(p, 0);
850 page_zone(page)->managed_pages += nr_pages;
851 set_page_refcounted(page);
852 __free_pages(page, order);
856 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
857 void __init init_cma_reserved_pageblock(struct page *page)
859 unsigned i = pageblock_nr_pages;
860 struct page *p = page;
863 __ClearPageReserved(p);
864 set_page_count(p, 0);
867 set_pageblock_migratetype(page, MIGRATE_CMA);
869 if (pageblock_order >= MAX_ORDER) {
870 i = pageblock_nr_pages;
873 set_page_refcounted(p);
874 __free_pages(p, MAX_ORDER - 1);
875 p += MAX_ORDER_NR_PAGES;
876 } while (i -= MAX_ORDER_NR_PAGES);
878 set_page_refcounted(page);
879 __free_pages(page, pageblock_order);
882 adjust_managed_page_count(page, pageblock_nr_pages);
887 * The order of subdivision here is critical for the IO subsystem.
888 * Please do not alter this order without good reasons and regression
889 * testing. Specifically, as large blocks of memory are subdivided,
890 * the order in which smaller blocks are delivered depends on the order
891 * they're subdivided in this function. This is the primary factor
892 * influencing the order in which pages are delivered to the IO
893 * subsystem according to empirical testing, and this is also justified
894 * by considering the behavior of a buddy system containing a single
895 * large block of memory acted on by a series of small allocations.
896 * This behavior is a critical factor in sglist merging's success.
900 static inline void expand(struct zone *zone, struct page *page,
901 int low, int high, struct free_area *area,
904 unsigned long size = 1 << high;
910 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
912 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
913 debug_guardpage_enabled() &&
914 high < debug_guardpage_minorder()) {
916 * Mark as guard pages (or page), that will allow to
917 * merge back to allocator when buddy will be freed.
918 * Corresponding page table entries will not be touched,
919 * pages will stay not present in virtual address space
921 set_page_guard(zone, &page[size], high, migratetype);
924 list_add(&page[size].lru, &area->free_list[migratetype]);
926 set_page_order(&page[size], high);
931 * This page is about to be returned from the page allocator
933 static inline int check_new_page(struct page *page)
935 const char *bad_reason = NULL;
936 unsigned long bad_flags = 0;
938 if (unlikely(page_mapcount(page)))
939 bad_reason = "nonzero mapcount";
940 if (unlikely(page->mapping != NULL))
941 bad_reason = "non-NULL mapping";
942 if (unlikely(atomic_read(&page->_count) != 0))
943 bad_reason = "nonzero _count";
944 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
945 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
946 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
949 if (unlikely(page->mem_cgroup))
950 bad_reason = "page still charged to cgroup";
952 if (unlikely(bad_reason)) {
953 bad_page(page, bad_reason, bad_flags);
959 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
964 for (i = 0; i < (1 << order); i++) {
965 struct page *p = page + i;
966 if (unlikely(check_new_page(p)))
970 set_page_private(page, 0);
971 set_page_refcounted(page);
973 arch_alloc_page(page, order);
974 kernel_map_pages(page, 1 << order, 1);
975 kasan_alloc_pages(page, order);
977 if (gfp_flags & __GFP_ZERO)
978 prep_zero_page(page, order, gfp_flags);
980 if (order && (gfp_flags & __GFP_COMP))
981 prep_compound_page(page, order);
983 set_page_owner(page, order, gfp_flags);
986 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
987 * allocate the page. The expectation is that the caller is taking
988 * steps that will free more memory. The caller should avoid the page
989 * being used for !PFMEMALLOC purposes.
991 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
997 * Go through the free lists for the given migratetype and remove
998 * the smallest available page from the freelists
1001 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1004 unsigned int current_order;
1005 struct free_area *area;
1008 /* Find a page of the appropriate size in the preferred list */
1009 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1010 area = &(zone->free_area[current_order]);
1011 if (list_empty(&area->free_list[migratetype]))
1014 page = list_entry(area->free_list[migratetype].next,
1016 list_del(&page->lru);
1017 rmv_page_order(page);
1019 expand(zone, page, order, current_order, area, migratetype);
1020 set_freepage_migratetype(page, migratetype);
1029 * This array describes the order lists are fallen back to when
1030 * the free lists for the desirable migrate type are depleted
1032 static int fallbacks[MIGRATE_TYPES][4] = {
1033 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1034 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1035 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1037 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1039 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1040 #ifdef CONFIG_MEMORY_ISOLATION
1041 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1046 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1049 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1052 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1053 unsigned int order) { return NULL; }
1057 * Move the free pages in a range to the free lists of the requested type.
1058 * Note that start_page and end_pages are not aligned on a pageblock
1059 * boundary. If alignment is required, use move_freepages_block()
1061 int move_freepages(struct zone *zone,
1062 struct page *start_page, struct page *end_page,
1066 unsigned long order;
1067 int pages_moved = 0;
1069 #ifndef CONFIG_HOLES_IN_ZONE
1071 * page_zone is not safe to call in this context when
1072 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1073 * anyway as we check zone boundaries in move_freepages_block().
1074 * Remove at a later date when no bug reports exist related to
1075 * grouping pages by mobility
1077 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1080 for (page = start_page; page <= end_page;) {
1081 /* Make sure we are not inadvertently changing nodes */
1082 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1084 if (!pfn_valid_within(page_to_pfn(page))) {
1089 if (!PageBuddy(page)) {
1094 order = page_order(page);
1095 list_move(&page->lru,
1096 &zone->free_area[order].free_list[migratetype]);
1097 set_freepage_migratetype(page, migratetype);
1099 pages_moved += 1 << order;
1105 int move_freepages_block(struct zone *zone, struct page *page,
1108 unsigned long start_pfn, end_pfn;
1109 struct page *start_page, *end_page;
1111 start_pfn = page_to_pfn(page);
1112 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1113 start_page = pfn_to_page(start_pfn);
1114 end_page = start_page + pageblock_nr_pages - 1;
1115 end_pfn = start_pfn + pageblock_nr_pages - 1;
1117 /* Do not cross zone boundaries */
1118 if (!zone_spans_pfn(zone, start_pfn))
1120 if (!zone_spans_pfn(zone, end_pfn))
1123 return move_freepages(zone, start_page, end_page, migratetype);
1126 static void change_pageblock_range(struct page *pageblock_page,
1127 int start_order, int migratetype)
1129 int nr_pageblocks = 1 << (start_order - pageblock_order);
1131 while (nr_pageblocks--) {
1132 set_pageblock_migratetype(pageblock_page, migratetype);
1133 pageblock_page += pageblock_nr_pages;
1138 * When we are falling back to another migratetype during allocation, try to
1139 * steal extra free pages from the same pageblocks to satisfy further
1140 * allocations, instead of polluting multiple pageblocks.
1142 * If we are stealing a relatively large buddy page, it is likely there will
1143 * be more free pages in the pageblock, so try to steal them all. For
1144 * reclaimable and unmovable allocations, we steal regardless of page size,
1145 * as fragmentation caused by those allocations polluting movable pageblocks
1146 * is worse than movable allocations stealing from unmovable and reclaimable
1149 static bool can_steal_fallback(unsigned int order, int start_mt)
1152 * Leaving this order check is intended, although there is
1153 * relaxed order check in next check. The reason is that
1154 * we can actually steal whole pageblock if this condition met,
1155 * but, below check doesn't guarantee it and that is just heuristic
1156 * so could be changed anytime.
1158 if (order >= pageblock_order)
1161 if (order >= pageblock_order / 2 ||
1162 start_mt == MIGRATE_RECLAIMABLE ||
1163 start_mt == MIGRATE_UNMOVABLE ||
1164 page_group_by_mobility_disabled)
1171 * This function implements actual steal behaviour. If order is large enough,
1172 * we can steal whole pageblock. If not, we first move freepages in this
1173 * pageblock and check whether half of pages are moved or not. If half of
1174 * pages are moved, we can change migratetype of pageblock and permanently
1175 * use it's pages as requested migratetype in the future.
1177 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1180 int current_order = page_order(page);
1183 /* Take ownership for orders >= pageblock_order */
1184 if (current_order >= pageblock_order) {
1185 change_pageblock_range(page, current_order, start_type);
1189 pages = move_freepages_block(zone, page, start_type);
1191 /* Claim the whole block if over half of it is free */
1192 if (pages >= (1 << (pageblock_order-1)) ||
1193 page_group_by_mobility_disabled)
1194 set_pageblock_migratetype(page, start_type);
1198 * Check whether there is a suitable fallback freepage with requested order.
1199 * If only_stealable is true, this function returns fallback_mt only if
1200 * we can steal other freepages all together. This would help to reduce
1201 * fragmentation due to mixed migratetype pages in one pageblock.
1203 int find_suitable_fallback(struct free_area *area, unsigned int order,
1204 int migratetype, bool only_stealable, bool *can_steal)
1209 if (area->nr_free == 0)
1214 fallback_mt = fallbacks[migratetype][i];
1215 if (fallback_mt == MIGRATE_RESERVE)
1218 if (list_empty(&area->free_list[fallback_mt]))
1221 if (can_steal_fallback(order, migratetype))
1224 if (!only_stealable)
1234 /* Remove an element from the buddy allocator from the fallback list */
1235 static inline struct page *
1236 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1238 struct free_area *area;
1239 unsigned int current_order;
1244 /* Find the largest possible block of pages in the other list */
1245 for (current_order = MAX_ORDER-1;
1246 current_order >= order && current_order <= MAX_ORDER-1;
1248 area = &(zone->free_area[current_order]);
1249 fallback_mt = find_suitable_fallback(area, current_order,
1250 start_migratetype, false, &can_steal);
1251 if (fallback_mt == -1)
1254 page = list_entry(area->free_list[fallback_mt].next,
1257 steal_suitable_fallback(zone, page, start_migratetype);
1259 /* Remove the page from the freelists */
1261 list_del(&page->lru);
1262 rmv_page_order(page);
1264 expand(zone, page, order, current_order, area,
1267 * The freepage_migratetype may differ from pageblock's
1268 * migratetype depending on the decisions in
1269 * try_to_steal_freepages(). This is OK as long as it
1270 * does not differ for MIGRATE_CMA pageblocks. For CMA
1271 * we need to make sure unallocated pages flushed from
1272 * pcp lists are returned to the correct freelist.
1274 set_freepage_migratetype(page, start_migratetype);
1276 trace_mm_page_alloc_extfrag(page, order, current_order,
1277 start_migratetype, fallback_mt);
1286 * Do the hard work of removing an element from the buddy allocator.
1287 * Call me with the zone->lock already held.
1289 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1295 page = __rmqueue_smallest(zone, order, migratetype);
1297 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1298 if (migratetype == MIGRATE_MOVABLE)
1299 page = __rmqueue_cma_fallback(zone, order);
1302 page = __rmqueue_fallback(zone, order, migratetype);
1305 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1306 * is used because __rmqueue_smallest is an inline function
1307 * and we want just one call site
1310 migratetype = MIGRATE_RESERVE;
1315 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1320 * Obtain a specified number of elements from the buddy allocator, all under
1321 * a single hold of the lock, for efficiency. Add them to the supplied list.
1322 * Returns the number of new pages which were placed at *list.
1324 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1325 unsigned long count, struct list_head *list,
1326 int migratetype, bool cold)
1330 spin_lock(&zone->lock);
1331 for (i = 0; i < count; ++i) {
1332 struct page *page = __rmqueue(zone, order, migratetype);
1333 if (unlikely(page == NULL))
1337 * Split buddy pages returned by expand() are received here
1338 * in physical page order. The page is added to the callers and
1339 * list and the list head then moves forward. From the callers
1340 * perspective, the linked list is ordered by page number in
1341 * some conditions. This is useful for IO devices that can
1342 * merge IO requests if the physical pages are ordered
1346 list_add(&page->lru, list);
1348 list_add_tail(&page->lru, list);
1350 if (is_migrate_cma(get_freepage_migratetype(page)))
1351 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1354 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1355 spin_unlock(&zone->lock);
1361 * Called from the vmstat counter updater to drain pagesets of this
1362 * currently executing processor on remote nodes after they have
1365 * Note that this function must be called with the thread pinned to
1366 * a single processor.
1368 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1370 unsigned long flags;
1371 int to_drain, batch;
1373 local_irq_save(flags);
1374 batch = READ_ONCE(pcp->batch);
1375 to_drain = min(pcp->count, batch);
1377 free_pcppages_bulk(zone, to_drain, pcp);
1378 pcp->count -= to_drain;
1380 local_irq_restore(flags);
1385 * Drain pcplists of the indicated processor and zone.
1387 * The processor must either be the current processor and the
1388 * thread pinned to the current processor or a processor that
1391 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1393 unsigned long flags;
1394 struct per_cpu_pageset *pset;
1395 struct per_cpu_pages *pcp;
1397 local_irq_save(flags);
1398 pset = per_cpu_ptr(zone->pageset, cpu);
1402 free_pcppages_bulk(zone, pcp->count, pcp);
1405 local_irq_restore(flags);
1409 * Drain pcplists of all zones on the indicated processor.
1411 * The processor must either be the current processor and the
1412 * thread pinned to the current processor or a processor that
1415 static void drain_pages(unsigned int cpu)
1419 for_each_populated_zone(zone) {
1420 drain_pages_zone(cpu, zone);
1425 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1427 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1428 * the single zone's pages.
1430 void drain_local_pages(struct zone *zone)
1432 int cpu = smp_processor_id();
1435 drain_pages_zone(cpu, zone);
1441 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1443 * When zone parameter is non-NULL, spill just the single zone's pages.
1445 * Note that this code is protected against sending an IPI to an offline
1446 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1447 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1448 * nothing keeps CPUs from showing up after we populated the cpumask and
1449 * before the call to on_each_cpu_mask().
1451 void drain_all_pages(struct zone *zone)
1456 * Allocate in the BSS so we wont require allocation in
1457 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1459 static cpumask_t cpus_with_pcps;
1462 * We don't care about racing with CPU hotplug event
1463 * as offline notification will cause the notified
1464 * cpu to drain that CPU pcps and on_each_cpu_mask
1465 * disables preemption as part of its processing
1467 for_each_online_cpu(cpu) {
1468 struct per_cpu_pageset *pcp;
1470 bool has_pcps = false;
1473 pcp = per_cpu_ptr(zone->pageset, cpu);
1477 for_each_populated_zone(z) {
1478 pcp = per_cpu_ptr(z->pageset, cpu);
1479 if (pcp->pcp.count) {
1487 cpumask_set_cpu(cpu, &cpus_with_pcps);
1489 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1491 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1495 #ifdef CONFIG_HIBERNATION
1497 void mark_free_pages(struct zone *zone)
1499 unsigned long pfn, max_zone_pfn;
1500 unsigned long flags;
1501 unsigned int order, t;
1502 struct list_head *curr;
1504 if (zone_is_empty(zone))
1507 spin_lock_irqsave(&zone->lock, flags);
1509 max_zone_pfn = zone_end_pfn(zone);
1510 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1511 if (pfn_valid(pfn)) {
1512 struct page *page = pfn_to_page(pfn);
1514 if (!swsusp_page_is_forbidden(page))
1515 swsusp_unset_page_free(page);
1518 for_each_migratetype_order(order, t) {
1519 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1522 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1523 for (i = 0; i < (1UL << order); i++)
1524 swsusp_set_page_free(pfn_to_page(pfn + i));
1527 spin_unlock_irqrestore(&zone->lock, flags);
1529 #endif /* CONFIG_PM */
1532 * Free a 0-order page
1533 * cold == true ? free a cold page : free a hot page
1535 void free_hot_cold_page(struct page *page, bool cold)
1537 struct zone *zone = page_zone(page);
1538 struct per_cpu_pages *pcp;
1539 unsigned long flags;
1540 unsigned long pfn = page_to_pfn(page);
1543 if (!free_pages_prepare(page, 0))
1546 migratetype = get_pfnblock_migratetype(page, pfn);
1547 set_freepage_migratetype(page, migratetype);
1548 local_irq_save(flags);
1549 __count_vm_event(PGFREE);
1552 * We only track unmovable, reclaimable and movable on pcp lists.
1553 * Free ISOLATE pages back to the allocator because they are being
1554 * offlined but treat RESERVE as movable pages so we can get those
1555 * areas back if necessary. Otherwise, we may have to free
1556 * excessively into the page allocator
1558 if (migratetype >= MIGRATE_PCPTYPES) {
1559 if (unlikely(is_migrate_isolate(migratetype))) {
1560 free_one_page(zone, page, pfn, 0, migratetype);
1563 migratetype = MIGRATE_MOVABLE;
1566 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1568 list_add(&page->lru, &pcp->lists[migratetype]);
1570 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1572 if (pcp->count >= pcp->high) {
1573 unsigned long batch = READ_ONCE(pcp->batch);
1574 free_pcppages_bulk(zone, batch, pcp);
1575 pcp->count -= batch;
1579 local_irq_restore(flags);
1583 * Free a list of 0-order pages
1585 void free_hot_cold_page_list(struct list_head *list, bool cold)
1587 struct page *page, *next;
1589 list_for_each_entry_safe(page, next, list, lru) {
1590 trace_mm_page_free_batched(page, cold);
1591 free_hot_cold_page(page, cold);
1596 * split_page takes a non-compound higher-order page, and splits it into
1597 * n (1<<order) sub-pages: page[0..n]
1598 * Each sub-page must be freed individually.
1600 * Note: this is probably too low level an operation for use in drivers.
1601 * Please consult with lkml before using this in your driver.
1603 void split_page(struct page *page, unsigned int order)
1607 VM_BUG_ON_PAGE(PageCompound(page), page);
1608 VM_BUG_ON_PAGE(!page_count(page), page);
1610 #ifdef CONFIG_KMEMCHECK
1612 * Split shadow pages too, because free(page[0]) would
1613 * otherwise free the whole shadow.
1615 if (kmemcheck_page_is_tracked(page))
1616 split_page(virt_to_page(page[0].shadow), order);
1619 set_page_owner(page, 0, 0);
1620 for (i = 1; i < (1 << order); i++) {
1621 set_page_refcounted(page + i);
1622 set_page_owner(page + i, 0, 0);
1625 EXPORT_SYMBOL_GPL(split_page);
1627 int __isolate_free_page(struct page *page, unsigned int order)
1629 unsigned long watermark;
1633 BUG_ON(!PageBuddy(page));
1635 zone = page_zone(page);
1636 mt = get_pageblock_migratetype(page);
1638 if (!is_migrate_isolate(mt)) {
1639 /* Obey watermarks as if the page was being allocated */
1640 watermark = low_wmark_pages(zone) + (1 << order);
1641 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1644 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1647 /* Remove page from free list */
1648 list_del(&page->lru);
1649 zone->free_area[order].nr_free--;
1650 rmv_page_order(page);
1652 /* Set the pageblock if the isolated page is at least a pageblock */
1653 if (order >= pageblock_order - 1) {
1654 struct page *endpage = page + (1 << order) - 1;
1655 for (; page < endpage; page += pageblock_nr_pages) {
1656 int mt = get_pageblock_migratetype(page);
1657 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1658 set_pageblock_migratetype(page,
1663 set_page_owner(page, order, 0);
1664 return 1UL << order;
1668 * Similar to split_page except the page is already free. As this is only
1669 * being used for migration, the migratetype of the block also changes.
1670 * As this is called with interrupts disabled, the caller is responsible
1671 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1674 * Note: this is probably too low level an operation for use in drivers.
1675 * Please consult with lkml before using this in your driver.
1677 int split_free_page(struct page *page)
1682 order = page_order(page);
1684 nr_pages = __isolate_free_page(page, order);
1688 /* Split into individual pages */
1689 set_page_refcounted(page);
1690 split_page(page, order);
1695 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1698 struct page *buffered_rmqueue(struct zone *preferred_zone,
1699 struct zone *zone, unsigned int order,
1700 gfp_t gfp_flags, int migratetype)
1702 unsigned long flags;
1704 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1706 if (likely(order == 0)) {
1707 struct per_cpu_pages *pcp;
1708 struct list_head *list;
1710 local_irq_save(flags);
1711 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1712 list = &pcp->lists[migratetype];
1713 if (list_empty(list)) {
1714 pcp->count += rmqueue_bulk(zone, 0,
1717 if (unlikely(list_empty(list)))
1722 page = list_entry(list->prev, struct page, lru);
1724 page = list_entry(list->next, struct page, lru);
1726 list_del(&page->lru);
1729 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1731 * __GFP_NOFAIL is not to be used in new code.
1733 * All __GFP_NOFAIL callers should be fixed so that they
1734 * properly detect and handle allocation failures.
1736 * We most definitely don't want callers attempting to
1737 * allocate greater than order-1 page units with
1740 WARN_ON_ONCE(order > 1);
1742 spin_lock_irqsave(&zone->lock, flags);
1743 page = __rmqueue(zone, order, migratetype);
1744 spin_unlock(&zone->lock);
1747 __mod_zone_freepage_state(zone, -(1 << order),
1748 get_freepage_migratetype(page));
1751 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1752 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1753 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1754 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1756 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1757 zone_statistics(preferred_zone, zone, gfp_flags);
1758 local_irq_restore(flags);
1760 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1764 local_irq_restore(flags);
1768 #ifdef CONFIG_FAIL_PAGE_ALLOC
1771 struct fault_attr attr;
1773 u32 ignore_gfp_highmem;
1774 u32 ignore_gfp_wait;
1776 } fail_page_alloc = {
1777 .attr = FAULT_ATTR_INITIALIZER,
1778 .ignore_gfp_wait = 1,
1779 .ignore_gfp_highmem = 1,
1783 static int __init setup_fail_page_alloc(char *str)
1785 return setup_fault_attr(&fail_page_alloc.attr, str);
1787 __setup("fail_page_alloc=", setup_fail_page_alloc);
1789 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1791 if (order < fail_page_alloc.min_order)
1793 if (gfp_mask & __GFP_NOFAIL)
1795 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1797 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1800 return should_fail(&fail_page_alloc.attr, 1 << order);
1803 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1805 static int __init fail_page_alloc_debugfs(void)
1807 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1810 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1811 &fail_page_alloc.attr);
1813 return PTR_ERR(dir);
1815 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1816 &fail_page_alloc.ignore_gfp_wait))
1818 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1819 &fail_page_alloc.ignore_gfp_highmem))
1821 if (!debugfs_create_u32("min-order", mode, dir,
1822 &fail_page_alloc.min_order))
1827 debugfs_remove_recursive(dir);
1832 late_initcall(fail_page_alloc_debugfs);
1834 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1836 #else /* CONFIG_FAIL_PAGE_ALLOC */
1838 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1843 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1846 * Return true if free pages are above 'mark'. This takes into account the order
1847 * of the allocation.
1849 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1850 unsigned long mark, int classzone_idx, int alloc_flags,
1853 /* free_pages may go negative - that's OK */
1858 free_pages -= (1 << order) - 1;
1859 if (alloc_flags & ALLOC_HIGH)
1861 if (alloc_flags & ALLOC_HARDER)
1864 /* If allocation can't use CMA areas don't use free CMA pages */
1865 if (!(alloc_flags & ALLOC_CMA))
1866 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1869 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1871 for (o = 0; o < order; o++) {
1872 /* At the next order, this order's pages become unavailable */
1873 free_pages -= z->free_area[o].nr_free << o;
1875 /* Require fewer higher order pages to be free */
1878 if (free_pages <= min)
1884 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1885 int classzone_idx, int alloc_flags)
1887 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1888 zone_page_state(z, NR_FREE_PAGES));
1891 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1892 unsigned long mark, int classzone_idx, int alloc_flags)
1894 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1896 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1897 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1899 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1905 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1906 * skip over zones that are not allowed by the cpuset, or that have
1907 * been recently (in last second) found to be nearly full. See further
1908 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1909 * that have to skip over a lot of full or unallowed zones.
1911 * If the zonelist cache is present in the passed zonelist, then
1912 * returns a pointer to the allowed node mask (either the current
1913 * tasks mems_allowed, or node_states[N_MEMORY].)
1915 * If the zonelist cache is not available for this zonelist, does
1916 * nothing and returns NULL.
1918 * If the fullzones BITMAP in the zonelist cache is stale (more than
1919 * a second since last zap'd) then we zap it out (clear its bits.)
1921 * We hold off even calling zlc_setup, until after we've checked the
1922 * first zone in the zonelist, on the theory that most allocations will
1923 * be satisfied from that first zone, so best to examine that zone as
1924 * quickly as we can.
1926 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1928 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1929 nodemask_t *allowednodes; /* zonelist_cache approximation */
1931 zlc = zonelist->zlcache_ptr;
1935 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1936 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1937 zlc->last_full_zap = jiffies;
1940 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1941 &cpuset_current_mems_allowed :
1942 &node_states[N_MEMORY];
1943 return allowednodes;
1947 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1948 * if it is worth looking at further for free memory:
1949 * 1) Check that the zone isn't thought to be full (doesn't have its
1950 * bit set in the zonelist_cache fullzones BITMAP).
1951 * 2) Check that the zones node (obtained from the zonelist_cache
1952 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1953 * Return true (non-zero) if zone is worth looking at further, or
1954 * else return false (zero) if it is not.
1956 * This check -ignores- the distinction between various watermarks,
1957 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1958 * found to be full for any variation of these watermarks, it will
1959 * be considered full for up to one second by all requests, unless
1960 * we are so low on memory on all allowed nodes that we are forced
1961 * into the second scan of the zonelist.
1963 * In the second scan we ignore this zonelist cache and exactly
1964 * apply the watermarks to all zones, even it is slower to do so.
1965 * We are low on memory in the second scan, and should leave no stone
1966 * unturned looking for a free page.
1968 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1969 nodemask_t *allowednodes)
1971 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1972 int i; /* index of *z in zonelist zones */
1973 int n; /* node that zone *z is on */
1975 zlc = zonelist->zlcache_ptr;
1979 i = z - zonelist->_zonerefs;
1982 /* This zone is worth trying if it is allowed but not full */
1983 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1987 * Given 'z' scanning a zonelist, set the corresponding bit in
1988 * zlc->fullzones, so that subsequent attempts to allocate a page
1989 * from that zone don't waste time re-examining it.
1991 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1993 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1994 int i; /* index of *z in zonelist zones */
1996 zlc = zonelist->zlcache_ptr;
2000 i = z - zonelist->_zonerefs;
2002 set_bit(i, zlc->fullzones);
2006 * clear all zones full, called after direct reclaim makes progress so that
2007 * a zone that was recently full is not skipped over for up to a second
2009 static void zlc_clear_zones_full(struct zonelist *zonelist)
2011 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2013 zlc = zonelist->zlcache_ptr;
2017 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2020 static bool zone_local(struct zone *local_zone, struct zone *zone)
2022 return local_zone->node == zone->node;
2025 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2027 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2031 #else /* CONFIG_NUMA */
2033 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2038 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2039 nodemask_t *allowednodes)
2044 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2048 static void zlc_clear_zones_full(struct zonelist *zonelist)
2052 static bool zone_local(struct zone *local_zone, struct zone *zone)
2057 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2062 #endif /* CONFIG_NUMA */
2064 static void reset_alloc_batches(struct zone *preferred_zone)
2066 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2069 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2070 high_wmark_pages(zone) - low_wmark_pages(zone) -
2071 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2072 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2073 } while (zone++ != preferred_zone);
2077 * get_page_from_freelist goes through the zonelist trying to allocate
2080 static struct page *
2081 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2082 const struct alloc_context *ac)
2084 struct zonelist *zonelist = ac->zonelist;
2086 struct page *page = NULL;
2088 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2089 int zlc_active = 0; /* set if using zonelist_cache */
2090 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2091 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2092 (gfp_mask & __GFP_WRITE);
2093 int nr_fair_skipped = 0;
2094 bool zonelist_rescan;
2097 zonelist_rescan = false;
2100 * Scan zonelist, looking for a zone with enough free.
2101 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2103 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2107 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2108 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2110 if (cpusets_enabled() &&
2111 (alloc_flags & ALLOC_CPUSET) &&
2112 !cpuset_zone_allowed(zone, gfp_mask))
2115 * Distribute pages in proportion to the individual
2116 * zone size to ensure fair page aging. The zone a
2117 * page was allocated in should have no effect on the
2118 * time the page has in memory before being reclaimed.
2120 if (alloc_flags & ALLOC_FAIR) {
2121 if (!zone_local(ac->preferred_zone, zone))
2123 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2129 * When allocating a page cache page for writing, we
2130 * want to get it from a zone that is within its dirty
2131 * limit, such that no single zone holds more than its
2132 * proportional share of globally allowed dirty pages.
2133 * The dirty limits take into account the zone's
2134 * lowmem reserves and high watermark so that kswapd
2135 * should be able to balance it without having to
2136 * write pages from its LRU list.
2138 * This may look like it could increase pressure on
2139 * lower zones by failing allocations in higher zones
2140 * before they are full. But the pages that do spill
2141 * over are limited as the lower zones are protected
2142 * by this very same mechanism. It should not become
2143 * a practical burden to them.
2145 * XXX: For now, allow allocations to potentially
2146 * exceed the per-zone dirty limit in the slowpath
2147 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2148 * which is important when on a NUMA setup the allowed
2149 * zones are together not big enough to reach the
2150 * global limit. The proper fix for these situations
2151 * will require awareness of zones in the
2152 * dirty-throttling and the flusher threads.
2154 if (consider_zone_dirty && !zone_dirty_ok(zone))
2157 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2158 if (!zone_watermark_ok(zone, order, mark,
2159 ac->classzone_idx, alloc_flags)) {
2162 /* Checked here to keep the fast path fast */
2163 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2164 if (alloc_flags & ALLOC_NO_WATERMARKS)
2167 if (IS_ENABLED(CONFIG_NUMA) &&
2168 !did_zlc_setup && nr_online_nodes > 1) {
2170 * we do zlc_setup if there are multiple nodes
2171 * and before considering the first zone allowed
2174 allowednodes = zlc_setup(zonelist, alloc_flags);
2179 if (zone_reclaim_mode == 0 ||
2180 !zone_allows_reclaim(ac->preferred_zone, zone))
2181 goto this_zone_full;
2184 * As we may have just activated ZLC, check if the first
2185 * eligible zone has failed zone_reclaim recently.
2187 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2188 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2191 ret = zone_reclaim(zone, gfp_mask, order);
2193 case ZONE_RECLAIM_NOSCAN:
2196 case ZONE_RECLAIM_FULL:
2197 /* scanned but unreclaimable */
2200 /* did we reclaim enough */
2201 if (zone_watermark_ok(zone, order, mark,
2202 ac->classzone_idx, alloc_flags))
2206 * Failed to reclaim enough to meet watermark.
2207 * Only mark the zone full if checking the min
2208 * watermark or if we failed to reclaim just
2209 * 1<<order pages or else the page allocator
2210 * fastpath will prematurely mark zones full
2211 * when the watermark is between the low and
2214 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2215 ret == ZONE_RECLAIM_SOME)
2216 goto this_zone_full;
2223 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2224 gfp_mask, ac->migratetype);
2226 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2231 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2232 zlc_mark_zone_full(zonelist, z);
2236 * The first pass makes sure allocations are spread fairly within the
2237 * local node. However, the local node might have free pages left
2238 * after the fairness batches are exhausted, and remote zones haven't
2239 * even been considered yet. Try once more without fairness, and
2240 * include remote zones now, before entering the slowpath and waking
2241 * kswapd: prefer spilling to a remote zone over swapping locally.
2243 if (alloc_flags & ALLOC_FAIR) {
2244 alloc_flags &= ~ALLOC_FAIR;
2245 if (nr_fair_skipped) {
2246 zonelist_rescan = true;
2247 reset_alloc_batches(ac->preferred_zone);
2249 if (nr_online_nodes > 1)
2250 zonelist_rescan = true;
2253 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2254 /* Disable zlc cache for second zonelist scan */
2256 zonelist_rescan = true;
2259 if (zonelist_rescan)
2266 * Large machines with many possible nodes should not always dump per-node
2267 * meminfo in irq context.
2269 static inline bool should_suppress_show_mem(void)
2274 ret = in_interrupt();
2279 static DEFINE_RATELIMIT_STATE(nopage_rs,
2280 DEFAULT_RATELIMIT_INTERVAL,
2281 DEFAULT_RATELIMIT_BURST);
2283 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2285 unsigned int filter = SHOW_MEM_FILTER_NODES;
2287 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2288 debug_guardpage_minorder() > 0)
2292 * This documents exceptions given to allocations in certain
2293 * contexts that are allowed to allocate outside current's set
2296 if (!(gfp_mask & __GFP_NOMEMALLOC))
2297 if (test_thread_flag(TIF_MEMDIE) ||
2298 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2299 filter &= ~SHOW_MEM_FILTER_NODES;
2300 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2301 filter &= ~SHOW_MEM_FILTER_NODES;
2304 struct va_format vaf;
2307 va_start(args, fmt);
2312 pr_warn("%pV", &vaf);
2317 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2318 current->comm, order, gfp_mask);
2321 if (!should_suppress_show_mem())
2326 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2327 unsigned long did_some_progress,
2328 unsigned long pages_reclaimed)
2330 /* Do not loop if specifically requested */
2331 if (gfp_mask & __GFP_NORETRY)
2334 /* Always retry if specifically requested */
2335 if (gfp_mask & __GFP_NOFAIL)
2339 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2340 * making forward progress without invoking OOM. Suspend also disables
2341 * storage devices so kswapd will not help. Bail if we are suspending.
2343 if (!did_some_progress && pm_suspended_storage())
2347 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2348 * means __GFP_NOFAIL, but that may not be true in other
2351 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2355 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2356 * specified, then we retry until we no longer reclaim any pages
2357 * (above), or we've reclaimed an order of pages at least as
2358 * large as the allocation's order. In both cases, if the
2359 * allocation still fails, we stop retrying.
2361 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2367 static inline struct page *
2368 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2369 const struct alloc_context *ac, unsigned long *did_some_progress)
2373 *did_some_progress = 0;
2376 * Acquire the per-zone oom lock for each zone. If that
2377 * fails, somebody else is making progress for us.
2379 if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) {
2380 *did_some_progress = 1;
2381 schedule_timeout_uninterruptible(1);
2386 * Go through the zonelist yet one more time, keep very high watermark
2387 * here, this is only to catch a parallel oom killing, we must fail if
2388 * we're still under heavy pressure.
2390 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2391 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2395 if (!(gfp_mask & __GFP_NOFAIL)) {
2396 /* Coredumps can quickly deplete all memory reserves */
2397 if (current->flags & PF_DUMPCORE)
2399 /* The OOM killer will not help higher order allocs */
2400 if (order > PAGE_ALLOC_COSTLY_ORDER)
2402 /* The OOM killer does not needlessly kill tasks for lowmem */
2403 if (ac->high_zoneidx < ZONE_NORMAL)
2405 /* The OOM killer does not compensate for light reclaim */
2406 if (!(gfp_mask & __GFP_FS)) {
2408 * XXX: Page reclaim didn't yield anything,
2409 * and the OOM killer can't be invoked, but
2410 * keep looping as per should_alloc_retry().
2412 *did_some_progress = 1;
2415 /* The OOM killer may not free memory on a specific node */
2416 if (gfp_mask & __GFP_THISNODE)
2419 /* Exhausted what can be done so it's blamo time */
2420 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2421 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2422 *did_some_progress = 1;
2424 oom_zonelist_unlock(ac->zonelist, gfp_mask);
2428 #ifdef CONFIG_COMPACTION
2429 /* Try memory compaction for high-order allocations before reclaim */
2430 static struct page *
2431 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2432 int alloc_flags, const struct alloc_context *ac,
2433 enum migrate_mode mode, int *contended_compaction,
2434 bool *deferred_compaction)
2436 unsigned long compact_result;
2442 current->flags |= PF_MEMALLOC;
2443 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2444 mode, contended_compaction);
2445 current->flags &= ~PF_MEMALLOC;
2447 switch (compact_result) {
2448 case COMPACT_DEFERRED:
2449 *deferred_compaction = true;
2451 case COMPACT_SKIPPED:
2458 * At least in one zone compaction wasn't deferred or skipped, so let's
2459 * count a compaction stall
2461 count_vm_event(COMPACTSTALL);
2463 page = get_page_from_freelist(gfp_mask, order,
2464 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2467 struct zone *zone = page_zone(page);
2469 zone->compact_blockskip_flush = false;
2470 compaction_defer_reset(zone, order, true);
2471 count_vm_event(COMPACTSUCCESS);
2476 * It's bad if compaction run occurs and fails. The most likely reason
2477 * is that pages exist, but not enough to satisfy watermarks.
2479 count_vm_event(COMPACTFAIL);
2486 static inline struct page *
2487 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2488 int alloc_flags, const struct alloc_context *ac,
2489 enum migrate_mode mode, int *contended_compaction,
2490 bool *deferred_compaction)
2494 #endif /* CONFIG_COMPACTION */
2496 /* Perform direct synchronous page reclaim */
2498 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2499 const struct alloc_context *ac)
2501 struct reclaim_state reclaim_state;
2506 /* We now go into synchronous reclaim */
2507 cpuset_memory_pressure_bump();
2508 current->flags |= PF_MEMALLOC;
2509 lockdep_set_current_reclaim_state(gfp_mask);
2510 reclaim_state.reclaimed_slab = 0;
2511 current->reclaim_state = &reclaim_state;
2513 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2516 current->reclaim_state = NULL;
2517 lockdep_clear_current_reclaim_state();
2518 current->flags &= ~PF_MEMALLOC;
2525 /* The really slow allocator path where we enter direct reclaim */
2526 static inline struct page *
2527 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2528 int alloc_flags, const struct alloc_context *ac,
2529 unsigned long *did_some_progress)
2531 struct page *page = NULL;
2532 bool drained = false;
2534 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2535 if (unlikely(!(*did_some_progress)))
2538 /* After successful reclaim, reconsider all zones for allocation */
2539 if (IS_ENABLED(CONFIG_NUMA))
2540 zlc_clear_zones_full(ac->zonelist);
2543 page = get_page_from_freelist(gfp_mask, order,
2544 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2547 * If an allocation failed after direct reclaim, it could be because
2548 * pages are pinned on the per-cpu lists. Drain them and try again
2550 if (!page && !drained) {
2551 drain_all_pages(NULL);
2560 * This is called in the allocator slow-path if the allocation request is of
2561 * sufficient urgency to ignore watermarks and take other desperate measures
2563 static inline struct page *
2564 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2565 const struct alloc_context *ac)
2570 page = get_page_from_freelist(gfp_mask, order,
2571 ALLOC_NO_WATERMARKS, ac);
2573 if (!page && gfp_mask & __GFP_NOFAIL)
2574 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2576 } while (!page && (gfp_mask & __GFP_NOFAIL));
2581 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2586 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2587 ac->high_zoneidx, ac->nodemask)
2588 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2592 gfp_to_alloc_flags(gfp_t gfp_mask)
2594 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2595 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2597 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2598 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2601 * The caller may dip into page reserves a bit more if the caller
2602 * cannot run direct reclaim, or if the caller has realtime scheduling
2603 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2604 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2606 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2610 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2611 * if it can't schedule.
2613 if (!(gfp_mask & __GFP_NOMEMALLOC))
2614 alloc_flags |= ALLOC_HARDER;
2616 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2617 * comment for __cpuset_node_allowed().
2619 alloc_flags &= ~ALLOC_CPUSET;
2620 } else if (unlikely(rt_task(current)) && !in_interrupt())
2621 alloc_flags |= ALLOC_HARDER;
2623 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2624 if (gfp_mask & __GFP_MEMALLOC)
2625 alloc_flags |= ALLOC_NO_WATERMARKS;
2626 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2627 alloc_flags |= ALLOC_NO_WATERMARKS;
2628 else if (!in_interrupt() &&
2629 ((current->flags & PF_MEMALLOC) ||
2630 unlikely(test_thread_flag(TIF_MEMDIE))))
2631 alloc_flags |= ALLOC_NO_WATERMARKS;
2634 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2635 alloc_flags |= ALLOC_CMA;
2640 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2642 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2645 static inline struct page *
2646 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2647 struct alloc_context *ac)
2649 const gfp_t wait = gfp_mask & __GFP_WAIT;
2650 struct page *page = NULL;
2652 unsigned long pages_reclaimed = 0;
2653 unsigned long did_some_progress;
2654 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2655 bool deferred_compaction = false;
2656 int contended_compaction = COMPACT_CONTENDED_NONE;
2659 * In the slowpath, we sanity check order to avoid ever trying to
2660 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2661 * be using allocators in order of preference for an area that is
2664 if (order >= MAX_ORDER) {
2665 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2670 * If this allocation cannot block and it is for a specific node, then
2671 * fail early. There's no need to wakeup kswapd or retry for a
2672 * speculative node-specific allocation.
2674 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2678 if (!(gfp_mask & __GFP_NO_KSWAPD))
2679 wake_all_kswapds(order, ac);
2682 * OK, we're below the kswapd watermark and have kicked background
2683 * reclaim. Now things get more complex, so set up alloc_flags according
2684 * to how we want to proceed.
2686 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2689 * Find the true preferred zone if the allocation is unconstrained by
2692 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2693 struct zoneref *preferred_zoneref;
2694 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2695 ac->high_zoneidx, NULL, &ac->preferred_zone);
2696 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2699 /* This is the last chance, in general, before the goto nopage. */
2700 page = get_page_from_freelist(gfp_mask, order,
2701 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2705 /* Allocate without watermarks if the context allows */
2706 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2708 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2709 * the allocation is high priority and these type of
2710 * allocations are system rather than user orientated
2712 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2714 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2721 /* Atomic allocations - we can't balance anything */
2724 * All existing users of the deprecated __GFP_NOFAIL are
2725 * blockable, so warn of any new users that actually allow this
2726 * type of allocation to fail.
2728 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2732 /* Avoid recursion of direct reclaim */
2733 if (current->flags & PF_MEMALLOC)
2736 /* Avoid allocations with no watermarks from looping endlessly */
2737 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2741 * Try direct compaction. The first pass is asynchronous. Subsequent
2742 * attempts after direct reclaim are synchronous
2744 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2746 &contended_compaction,
2747 &deferred_compaction);
2751 /* Checks for THP-specific high-order allocations */
2752 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2754 * If compaction is deferred for high-order allocations, it is
2755 * because sync compaction recently failed. If this is the case
2756 * and the caller requested a THP allocation, we do not want
2757 * to heavily disrupt the system, so we fail the allocation
2758 * instead of entering direct reclaim.
2760 if (deferred_compaction)
2764 * In all zones where compaction was attempted (and not
2765 * deferred or skipped), lock contention has been detected.
2766 * For THP allocation we do not want to disrupt the others
2767 * so we fallback to base pages instead.
2769 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2773 * If compaction was aborted due to need_resched(), we do not
2774 * want to further increase allocation latency, unless it is
2775 * khugepaged trying to collapse.
2777 if (contended_compaction == COMPACT_CONTENDED_SCHED
2778 && !(current->flags & PF_KTHREAD))
2783 * It can become very expensive to allocate transparent hugepages at
2784 * fault, so use asynchronous memory compaction for THP unless it is
2785 * khugepaged trying to collapse.
2787 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2788 (current->flags & PF_KTHREAD))
2789 migration_mode = MIGRATE_SYNC_LIGHT;
2791 /* Try direct reclaim and then allocating */
2792 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2793 &did_some_progress);
2797 /* Check if we should retry the allocation */
2798 pages_reclaimed += did_some_progress;
2799 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2802 * If we fail to make progress by freeing individual
2803 * pages, but the allocation wants us to keep going,
2804 * start OOM killing tasks.
2806 if (!did_some_progress) {
2807 page = __alloc_pages_may_oom(gfp_mask, order, ac,
2808 &did_some_progress);
2811 if (!did_some_progress)
2814 /* Wait for some write requests to complete then retry */
2815 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2819 * High-order allocations do not necessarily loop after
2820 * direct reclaim and reclaim/compaction depends on compaction
2821 * being called after reclaim so call directly if necessary
2823 page = __alloc_pages_direct_compact(gfp_mask, order,
2824 alloc_flags, ac, migration_mode,
2825 &contended_compaction,
2826 &deferred_compaction);
2832 warn_alloc_failed(gfp_mask, order, NULL);
2838 * This is the 'heart' of the zoned buddy allocator.
2841 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2842 struct zonelist *zonelist, nodemask_t *nodemask)
2844 struct zoneref *preferred_zoneref;
2845 struct page *page = NULL;
2846 unsigned int cpuset_mems_cookie;
2847 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2848 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2849 struct alloc_context ac = {
2850 .high_zoneidx = gfp_zone(gfp_mask),
2851 .nodemask = nodemask,
2852 .migratetype = gfpflags_to_migratetype(gfp_mask),
2855 gfp_mask &= gfp_allowed_mask;
2857 lockdep_trace_alloc(gfp_mask);
2859 might_sleep_if(gfp_mask & __GFP_WAIT);
2861 if (should_fail_alloc_page(gfp_mask, order))
2865 * Check the zones suitable for the gfp_mask contain at least one
2866 * valid zone. It's possible to have an empty zonelist as a result
2867 * of __GFP_THISNODE and a memoryless node
2869 if (unlikely(!zonelist->_zonerefs->zone))
2872 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2873 alloc_flags |= ALLOC_CMA;
2876 cpuset_mems_cookie = read_mems_allowed_begin();
2878 /* We set it here, as __alloc_pages_slowpath might have changed it */
2879 ac.zonelist = zonelist;
2880 /* The preferred zone is used for statistics later */
2881 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2882 ac.nodemask ? : &cpuset_current_mems_allowed,
2883 &ac.preferred_zone);
2884 if (!ac.preferred_zone)
2886 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2888 /* First allocation attempt */
2889 alloc_mask = gfp_mask|__GFP_HARDWALL;
2890 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2891 if (unlikely(!page)) {
2893 * Runtime PM, block IO and its error handling path
2894 * can deadlock because I/O on the device might not
2897 alloc_mask = memalloc_noio_flags(gfp_mask);
2899 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2902 if (kmemcheck_enabled && page)
2903 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2905 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2909 * When updating a task's mems_allowed, it is possible to race with
2910 * parallel threads in such a way that an allocation can fail while
2911 * the mask is being updated. If a page allocation is about to fail,
2912 * check if the cpuset changed during allocation and if so, retry.
2914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2919 EXPORT_SYMBOL(__alloc_pages_nodemask);
2922 * Common helper functions.
2924 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2929 * __get_free_pages() returns a 32-bit address, which cannot represent
2932 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2934 page = alloc_pages(gfp_mask, order);
2937 return (unsigned long) page_address(page);
2939 EXPORT_SYMBOL(__get_free_pages);
2941 unsigned long get_zeroed_page(gfp_t gfp_mask)
2943 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2945 EXPORT_SYMBOL(get_zeroed_page);
2947 void __free_pages(struct page *page, unsigned int order)
2949 if (put_page_testzero(page)) {
2951 free_hot_cold_page(page, false);
2953 __free_pages_ok(page, order);
2957 EXPORT_SYMBOL(__free_pages);
2959 void free_pages(unsigned long addr, unsigned int order)
2962 VM_BUG_ON(!virt_addr_valid((void *)addr));
2963 __free_pages(virt_to_page((void *)addr), order);
2967 EXPORT_SYMBOL(free_pages);
2971 * An arbitrary-length arbitrary-offset area of memory which resides
2972 * within a 0 or higher order page. Multiple fragments within that page
2973 * are individually refcounted, in the page's reference counter.
2975 * The page_frag functions below provide a simple allocation framework for
2976 * page fragments. This is used by the network stack and network device
2977 * drivers to provide a backing region of memory for use as either an
2978 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
2980 static struct page *__page_frag_refill(struct page_frag_cache *nc,
2983 struct page *page = NULL;
2984 gfp_t gfp = gfp_mask;
2986 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
2987 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
2989 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
2990 PAGE_FRAG_CACHE_MAX_ORDER);
2991 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
2993 if (unlikely(!page))
2994 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
2996 nc->va = page ? page_address(page) : NULL;
3001 void *__alloc_page_frag(struct page_frag_cache *nc,
3002 unsigned int fragsz, gfp_t gfp_mask)
3004 unsigned int size = PAGE_SIZE;
3008 if (unlikely(!nc->va)) {
3010 page = __page_frag_refill(nc, gfp_mask);
3014 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3015 /* if size can vary use size else just use PAGE_SIZE */
3018 /* Even if we own the page, we do not use atomic_set().
3019 * This would break get_page_unless_zero() users.
3021 atomic_add(size - 1, &page->_count);
3023 /* reset page count bias and offset to start of new frag */
3024 nc->pfmemalloc = page->pfmemalloc;
3025 nc->pagecnt_bias = size;
3029 offset = nc->offset - fragsz;
3030 if (unlikely(offset < 0)) {
3031 page = virt_to_page(nc->va);
3033 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3036 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3037 /* if size can vary use size else just use PAGE_SIZE */
3040 /* OK, page count is 0, we can safely set it */
3041 atomic_set(&page->_count, size);
3043 /* reset page count bias and offset to start of new frag */
3044 nc->pagecnt_bias = size;
3045 offset = size - fragsz;
3049 nc->offset = offset;
3051 return nc->va + offset;
3053 EXPORT_SYMBOL(__alloc_page_frag);
3056 * Frees a page fragment allocated out of either a compound or order 0 page.
3058 void __free_page_frag(void *addr)
3060 struct page *page = virt_to_head_page(addr);
3062 if (unlikely(put_page_testzero(page)))
3063 __free_pages_ok(page, compound_order(page));
3065 EXPORT_SYMBOL(__free_page_frag);
3068 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3069 * of the current memory cgroup.
3071 * It should be used when the caller would like to use kmalloc, but since the
3072 * allocation is large, it has to fall back to the page allocator.
3074 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3077 struct mem_cgroup *memcg = NULL;
3079 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3081 page = alloc_pages(gfp_mask, order);
3082 memcg_kmem_commit_charge(page, memcg, order);
3086 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3089 struct mem_cgroup *memcg = NULL;
3091 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3093 page = alloc_pages_node(nid, gfp_mask, order);
3094 memcg_kmem_commit_charge(page, memcg, order);
3099 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3102 void __free_kmem_pages(struct page *page, unsigned int order)
3104 memcg_kmem_uncharge_pages(page, order);
3105 __free_pages(page, order);
3108 void free_kmem_pages(unsigned long addr, unsigned int order)
3111 VM_BUG_ON(!virt_addr_valid((void *)addr));
3112 __free_kmem_pages(virt_to_page((void *)addr), order);
3116 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3119 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3120 unsigned long used = addr + PAGE_ALIGN(size);
3122 split_page(virt_to_page((void *)addr), order);
3123 while (used < alloc_end) {
3128 return (void *)addr;
3132 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3133 * @size: the number of bytes to allocate
3134 * @gfp_mask: GFP flags for the allocation
3136 * This function is similar to alloc_pages(), except that it allocates the
3137 * minimum number of pages to satisfy the request. alloc_pages() can only
3138 * allocate memory in power-of-two pages.
3140 * This function is also limited by MAX_ORDER.
3142 * Memory allocated by this function must be released by free_pages_exact().
3144 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3146 unsigned int order = get_order(size);
3149 addr = __get_free_pages(gfp_mask, order);
3150 return make_alloc_exact(addr, order, size);
3152 EXPORT_SYMBOL(alloc_pages_exact);
3155 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3157 * @nid: the preferred node ID where memory should be allocated
3158 * @size: the number of bytes to allocate
3159 * @gfp_mask: GFP flags for the allocation
3161 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3163 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3166 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3168 unsigned order = get_order(size);
3169 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3172 return make_alloc_exact((unsigned long)page_address(p), order, size);
3176 * free_pages_exact - release memory allocated via alloc_pages_exact()
3177 * @virt: the value returned by alloc_pages_exact.
3178 * @size: size of allocation, same value as passed to alloc_pages_exact().
3180 * Release the memory allocated by a previous call to alloc_pages_exact.
3182 void free_pages_exact(void *virt, size_t size)
3184 unsigned long addr = (unsigned long)virt;
3185 unsigned long end = addr + PAGE_ALIGN(size);
3187 while (addr < end) {
3192 EXPORT_SYMBOL(free_pages_exact);
3195 * nr_free_zone_pages - count number of pages beyond high watermark
3196 * @offset: The zone index of the highest zone
3198 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3199 * high watermark within all zones at or below a given zone index. For each
3200 * zone, the number of pages is calculated as:
3201 * managed_pages - high_pages
3203 static unsigned long nr_free_zone_pages(int offset)
3208 /* Just pick one node, since fallback list is circular */
3209 unsigned long sum = 0;
3211 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3213 for_each_zone_zonelist(zone, z, zonelist, offset) {
3214 unsigned long size = zone->managed_pages;
3215 unsigned long high = high_wmark_pages(zone);
3224 * nr_free_buffer_pages - count number of pages beyond high watermark
3226 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3227 * watermark within ZONE_DMA and ZONE_NORMAL.
3229 unsigned long nr_free_buffer_pages(void)
3231 return nr_free_zone_pages(gfp_zone(GFP_USER));
3233 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3236 * nr_free_pagecache_pages - count number of pages beyond high watermark
3238 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3239 * high watermark within all zones.
3241 unsigned long nr_free_pagecache_pages(void)
3243 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3246 static inline void show_node(struct zone *zone)
3248 if (IS_ENABLED(CONFIG_NUMA))
3249 printk("Node %d ", zone_to_nid(zone));
3252 void si_meminfo(struct sysinfo *val)
3254 val->totalram = totalram_pages;
3255 val->sharedram = global_page_state(NR_SHMEM);
3256 val->freeram = global_page_state(NR_FREE_PAGES);
3257 val->bufferram = nr_blockdev_pages();
3258 val->totalhigh = totalhigh_pages;
3259 val->freehigh = nr_free_highpages();
3260 val->mem_unit = PAGE_SIZE;
3263 EXPORT_SYMBOL(si_meminfo);
3266 void si_meminfo_node(struct sysinfo *val, int nid)
3268 int zone_type; /* needs to be signed */
3269 unsigned long managed_pages = 0;
3270 pg_data_t *pgdat = NODE_DATA(nid);
3272 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3273 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3274 val->totalram = managed_pages;
3275 val->sharedram = node_page_state(nid, NR_SHMEM);
3276 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3277 #ifdef CONFIG_HIGHMEM
3278 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3279 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3285 val->mem_unit = PAGE_SIZE;
3290 * Determine whether the node should be displayed or not, depending on whether
3291 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3293 bool skip_free_areas_node(unsigned int flags, int nid)
3296 unsigned int cpuset_mems_cookie;
3298 if (!(flags & SHOW_MEM_FILTER_NODES))
3302 cpuset_mems_cookie = read_mems_allowed_begin();
3303 ret = !node_isset(nid, cpuset_current_mems_allowed);
3304 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3309 #define K(x) ((x) << (PAGE_SHIFT-10))
3311 static void show_migration_types(unsigned char type)
3313 static const char types[MIGRATE_TYPES] = {
3314 [MIGRATE_UNMOVABLE] = 'U',
3315 [MIGRATE_RECLAIMABLE] = 'E',
3316 [MIGRATE_MOVABLE] = 'M',
3317 [MIGRATE_RESERVE] = 'R',
3319 [MIGRATE_CMA] = 'C',
3321 #ifdef CONFIG_MEMORY_ISOLATION
3322 [MIGRATE_ISOLATE] = 'I',
3325 char tmp[MIGRATE_TYPES + 1];
3329 for (i = 0; i < MIGRATE_TYPES; i++) {
3330 if (type & (1 << i))
3335 printk("(%s) ", tmp);
3339 * Show free area list (used inside shift_scroll-lock stuff)
3340 * We also calculate the percentage fragmentation. We do this by counting the
3341 * memory on each free list with the exception of the first item on the list.
3344 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3347 void show_free_areas(unsigned int filter)
3349 unsigned long free_pcp = 0;
3353 for_each_populated_zone(zone) {
3354 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3357 for_each_online_cpu(cpu)
3358 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3361 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3362 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3363 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3364 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3365 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3366 " free:%lu free_pcp:%lu free_cma:%lu\n",
3367 global_page_state(NR_ACTIVE_ANON),
3368 global_page_state(NR_INACTIVE_ANON),
3369 global_page_state(NR_ISOLATED_ANON),
3370 global_page_state(NR_ACTIVE_FILE),
3371 global_page_state(NR_INACTIVE_FILE),
3372 global_page_state(NR_ISOLATED_FILE),
3373 global_page_state(NR_UNEVICTABLE),
3374 global_page_state(NR_FILE_DIRTY),
3375 global_page_state(NR_WRITEBACK),
3376 global_page_state(NR_UNSTABLE_NFS),
3377 global_page_state(NR_SLAB_RECLAIMABLE),
3378 global_page_state(NR_SLAB_UNRECLAIMABLE),
3379 global_page_state(NR_FILE_MAPPED),
3380 global_page_state(NR_SHMEM),
3381 global_page_state(NR_PAGETABLE),
3382 global_page_state(NR_BOUNCE),
3383 global_page_state(NR_FREE_PAGES),
3385 global_page_state(NR_FREE_CMA_PAGES));
3387 for_each_populated_zone(zone) {
3390 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3394 for_each_online_cpu(cpu)
3395 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3403 " active_anon:%lukB"
3404 " inactive_anon:%lukB"
3405 " active_file:%lukB"
3406 " inactive_file:%lukB"
3407 " unevictable:%lukB"
3408 " isolated(anon):%lukB"
3409 " isolated(file):%lukB"
3417 " slab_reclaimable:%lukB"
3418 " slab_unreclaimable:%lukB"
3419 " kernel_stack:%lukB"
3426 " writeback_tmp:%lukB"
3427 " pages_scanned:%lu"
3428 " all_unreclaimable? %s"
3431 K(zone_page_state(zone, NR_FREE_PAGES)),
3432 K(min_wmark_pages(zone)),
3433 K(low_wmark_pages(zone)),
3434 K(high_wmark_pages(zone)),
3435 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3436 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3437 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3438 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3439 K(zone_page_state(zone, NR_UNEVICTABLE)),
3440 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3441 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3442 K(zone->present_pages),
3443 K(zone->managed_pages),
3444 K(zone_page_state(zone, NR_MLOCK)),
3445 K(zone_page_state(zone, NR_FILE_DIRTY)),
3446 K(zone_page_state(zone, NR_WRITEBACK)),
3447 K(zone_page_state(zone, NR_FILE_MAPPED)),
3448 K(zone_page_state(zone, NR_SHMEM)),
3449 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3450 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3451 zone_page_state(zone, NR_KERNEL_STACK) *
3453 K(zone_page_state(zone, NR_PAGETABLE)),
3454 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3455 K(zone_page_state(zone, NR_BOUNCE)),
3457 K(this_cpu_read(zone->pageset->pcp.count)),
3458 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3459 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3460 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3461 (!zone_reclaimable(zone) ? "yes" : "no")
3463 printk("lowmem_reserve[]:");
3464 for (i = 0; i < MAX_NR_ZONES; i++)
3465 printk(" %ld", zone->lowmem_reserve[i]);
3469 for_each_populated_zone(zone) {
3470 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3471 unsigned char types[MAX_ORDER];
3473 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3476 printk("%s: ", zone->name);
3478 spin_lock_irqsave(&zone->lock, flags);
3479 for (order = 0; order < MAX_ORDER; order++) {
3480 struct free_area *area = &zone->free_area[order];
3483 nr[order] = area->nr_free;
3484 total += nr[order] << order;
3487 for (type = 0; type < MIGRATE_TYPES; type++) {
3488 if (!list_empty(&area->free_list[type]))
3489 types[order] |= 1 << type;
3492 spin_unlock_irqrestore(&zone->lock, flags);
3493 for (order = 0; order < MAX_ORDER; order++) {
3494 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3496 show_migration_types(types[order]);
3498 printk("= %lukB\n", K(total));
3501 hugetlb_show_meminfo();
3503 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3505 show_swap_cache_info();
3508 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3510 zoneref->zone = zone;
3511 zoneref->zone_idx = zone_idx(zone);
3515 * Builds allocation fallback zone lists.
3517 * Add all populated zones of a node to the zonelist.
3519 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3523 enum zone_type zone_type = MAX_NR_ZONES;
3527 zone = pgdat->node_zones + zone_type;
3528 if (populated_zone(zone)) {
3529 zoneref_set_zone(zone,
3530 &zonelist->_zonerefs[nr_zones++]);
3531 check_highest_zone(zone_type);
3533 } while (zone_type);
3541 * 0 = automatic detection of better ordering.
3542 * 1 = order by ([node] distance, -zonetype)
3543 * 2 = order by (-zonetype, [node] distance)
3545 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3546 * the same zonelist. So only NUMA can configure this param.
3548 #define ZONELIST_ORDER_DEFAULT 0
3549 #define ZONELIST_ORDER_NODE 1
3550 #define ZONELIST_ORDER_ZONE 2
3552 /* zonelist order in the kernel.
3553 * set_zonelist_order() will set this to NODE or ZONE.
3555 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3556 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3560 /* The value user specified ....changed by config */
3561 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3562 /* string for sysctl */
3563 #define NUMA_ZONELIST_ORDER_LEN 16
3564 char numa_zonelist_order[16] = "default";
3567 * interface for configure zonelist ordering.
3568 * command line option "numa_zonelist_order"
3569 * = "[dD]efault - default, automatic configuration.
3570 * = "[nN]ode - order by node locality, then by zone within node
3571 * = "[zZ]one - order by zone, then by locality within zone
3574 static int __parse_numa_zonelist_order(char *s)
3576 if (*s == 'd' || *s == 'D') {
3577 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3578 } else if (*s == 'n' || *s == 'N') {
3579 user_zonelist_order = ZONELIST_ORDER_NODE;
3580 } else if (*s == 'z' || *s == 'Z') {
3581 user_zonelist_order = ZONELIST_ORDER_ZONE;
3584 "Ignoring invalid numa_zonelist_order value: "
3591 static __init int setup_numa_zonelist_order(char *s)
3598 ret = __parse_numa_zonelist_order(s);
3600 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3604 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3607 * sysctl handler for numa_zonelist_order
3609 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3610 void __user *buffer, size_t *length,
3613 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3615 static DEFINE_MUTEX(zl_order_mutex);
3617 mutex_lock(&zl_order_mutex);
3619 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3623 strcpy(saved_string, (char *)table->data);
3625 ret = proc_dostring(table, write, buffer, length, ppos);
3629 int oldval = user_zonelist_order;
3631 ret = __parse_numa_zonelist_order((char *)table->data);
3634 * bogus value. restore saved string
3636 strncpy((char *)table->data, saved_string,
3637 NUMA_ZONELIST_ORDER_LEN);
3638 user_zonelist_order = oldval;
3639 } else if (oldval != user_zonelist_order) {
3640 mutex_lock(&zonelists_mutex);
3641 build_all_zonelists(NULL, NULL);
3642 mutex_unlock(&zonelists_mutex);
3646 mutex_unlock(&zl_order_mutex);
3651 #define MAX_NODE_LOAD (nr_online_nodes)
3652 static int node_load[MAX_NUMNODES];
3655 * find_next_best_node - find the next node that should appear in a given node's fallback list
3656 * @node: node whose fallback list we're appending
3657 * @used_node_mask: nodemask_t of already used nodes
3659 * We use a number of factors to determine which is the next node that should
3660 * appear on a given node's fallback list. The node should not have appeared
3661 * already in @node's fallback list, and it should be the next closest node
3662 * according to the distance array (which contains arbitrary distance values
3663 * from each node to each node in the system), and should also prefer nodes
3664 * with no CPUs, since presumably they'll have very little allocation pressure
3665 * on them otherwise.
3666 * It returns -1 if no node is found.
3668 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3671 int min_val = INT_MAX;
3672 int best_node = NUMA_NO_NODE;
3673 const struct cpumask *tmp = cpumask_of_node(0);
3675 /* Use the local node if we haven't already */
3676 if (!node_isset(node, *used_node_mask)) {
3677 node_set(node, *used_node_mask);
3681 for_each_node_state(n, N_MEMORY) {
3683 /* Don't want a node to appear more than once */
3684 if (node_isset(n, *used_node_mask))
3687 /* Use the distance array to find the distance */
3688 val = node_distance(node, n);
3690 /* Penalize nodes under us ("prefer the next node") */
3693 /* Give preference to headless and unused nodes */
3694 tmp = cpumask_of_node(n);
3695 if (!cpumask_empty(tmp))
3696 val += PENALTY_FOR_NODE_WITH_CPUS;
3698 /* Slight preference for less loaded node */
3699 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3700 val += node_load[n];
3702 if (val < min_val) {
3709 node_set(best_node, *used_node_mask);
3716 * Build zonelists ordered by node and zones within node.
3717 * This results in maximum locality--normal zone overflows into local
3718 * DMA zone, if any--but risks exhausting DMA zone.
3720 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3723 struct zonelist *zonelist;
3725 zonelist = &pgdat->node_zonelists[0];
3726 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3728 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3729 zonelist->_zonerefs[j].zone = NULL;
3730 zonelist->_zonerefs[j].zone_idx = 0;
3734 * Build gfp_thisnode zonelists
3736 static void build_thisnode_zonelists(pg_data_t *pgdat)
3739 struct zonelist *zonelist;
3741 zonelist = &pgdat->node_zonelists[1];
3742 j = build_zonelists_node(pgdat, zonelist, 0);
3743 zonelist->_zonerefs[j].zone = NULL;
3744 zonelist->_zonerefs[j].zone_idx = 0;
3748 * Build zonelists ordered by zone and nodes within zones.
3749 * This results in conserving DMA zone[s] until all Normal memory is
3750 * exhausted, but results in overflowing to remote node while memory
3751 * may still exist in local DMA zone.
3753 static int node_order[MAX_NUMNODES];
3755 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3758 int zone_type; /* needs to be signed */
3760 struct zonelist *zonelist;
3762 zonelist = &pgdat->node_zonelists[0];
3764 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3765 for (j = 0; j < nr_nodes; j++) {
3766 node = node_order[j];
3767 z = &NODE_DATA(node)->node_zones[zone_type];
3768 if (populated_zone(z)) {
3770 &zonelist->_zonerefs[pos++]);
3771 check_highest_zone(zone_type);
3775 zonelist->_zonerefs[pos].zone = NULL;
3776 zonelist->_zonerefs[pos].zone_idx = 0;
3779 #if defined(CONFIG_64BIT)
3781 * Devices that require DMA32/DMA are relatively rare and do not justify a
3782 * penalty to every machine in case the specialised case applies. Default
3783 * to Node-ordering on 64-bit NUMA machines
3785 static int default_zonelist_order(void)
3787 return ZONELIST_ORDER_NODE;
3791 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3792 * by the kernel. If processes running on node 0 deplete the low memory zone
3793 * then reclaim will occur more frequency increasing stalls and potentially
3794 * be easier to OOM if a large percentage of the zone is under writeback or
3795 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3796 * Hence, default to zone ordering on 32-bit.
3798 static int default_zonelist_order(void)
3800 return ZONELIST_ORDER_ZONE;
3802 #endif /* CONFIG_64BIT */
3804 static void set_zonelist_order(void)
3806 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3807 current_zonelist_order = default_zonelist_order();
3809 current_zonelist_order = user_zonelist_order;
3812 static void build_zonelists(pg_data_t *pgdat)
3816 nodemask_t used_mask;
3817 int local_node, prev_node;
3818 struct zonelist *zonelist;
3819 int order = current_zonelist_order;
3821 /* initialize zonelists */
3822 for (i = 0; i < MAX_ZONELISTS; i++) {
3823 zonelist = pgdat->node_zonelists + i;
3824 zonelist->_zonerefs[0].zone = NULL;
3825 zonelist->_zonerefs[0].zone_idx = 0;
3828 /* NUMA-aware ordering of nodes */
3829 local_node = pgdat->node_id;
3830 load = nr_online_nodes;
3831 prev_node = local_node;
3832 nodes_clear(used_mask);
3834 memset(node_order, 0, sizeof(node_order));
3837 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3839 * We don't want to pressure a particular node.
3840 * So adding penalty to the first node in same
3841 * distance group to make it round-robin.
3843 if (node_distance(local_node, node) !=
3844 node_distance(local_node, prev_node))
3845 node_load[node] = load;
3849 if (order == ZONELIST_ORDER_NODE)
3850 build_zonelists_in_node_order(pgdat, node);
3852 node_order[j++] = node; /* remember order */
3855 if (order == ZONELIST_ORDER_ZONE) {
3856 /* calculate node order -- i.e., DMA last! */
3857 build_zonelists_in_zone_order(pgdat, j);
3860 build_thisnode_zonelists(pgdat);
3863 /* Construct the zonelist performance cache - see further mmzone.h */
3864 static void build_zonelist_cache(pg_data_t *pgdat)
3866 struct zonelist *zonelist;
3867 struct zonelist_cache *zlc;
3870 zonelist = &pgdat->node_zonelists[0];
3871 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3872 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3873 for (z = zonelist->_zonerefs; z->zone; z++)
3874 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3877 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3879 * Return node id of node used for "local" allocations.
3880 * I.e., first node id of first zone in arg node's generic zonelist.
3881 * Used for initializing percpu 'numa_mem', which is used primarily
3882 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3884 int local_memory_node(int node)
3888 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3889 gfp_zone(GFP_KERNEL),
3896 #else /* CONFIG_NUMA */
3898 static void set_zonelist_order(void)
3900 current_zonelist_order = ZONELIST_ORDER_ZONE;
3903 static void build_zonelists(pg_data_t *pgdat)
3905 int node, local_node;
3907 struct zonelist *zonelist;
3909 local_node = pgdat->node_id;
3911 zonelist = &pgdat->node_zonelists[0];
3912 j = build_zonelists_node(pgdat, zonelist, 0);
3915 * Now we build the zonelist so that it contains the zones
3916 * of all the other nodes.
3917 * We don't want to pressure a particular node, so when
3918 * building the zones for node N, we make sure that the
3919 * zones coming right after the local ones are those from
3920 * node N+1 (modulo N)
3922 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3923 if (!node_online(node))
3925 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3927 for (node = 0; node < local_node; node++) {
3928 if (!node_online(node))
3930 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3933 zonelist->_zonerefs[j].zone = NULL;
3934 zonelist->_zonerefs[j].zone_idx = 0;
3937 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3938 static void build_zonelist_cache(pg_data_t *pgdat)
3940 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3943 #endif /* CONFIG_NUMA */
3946 * Boot pageset table. One per cpu which is going to be used for all
3947 * zones and all nodes. The parameters will be set in such a way
3948 * that an item put on a list will immediately be handed over to
3949 * the buddy list. This is safe since pageset manipulation is done
3950 * with interrupts disabled.
3952 * The boot_pagesets must be kept even after bootup is complete for
3953 * unused processors and/or zones. They do play a role for bootstrapping
3954 * hotplugged processors.
3956 * zoneinfo_show() and maybe other functions do
3957 * not check if the processor is online before following the pageset pointer.
3958 * Other parts of the kernel may not check if the zone is available.
3960 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3961 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3962 static void setup_zone_pageset(struct zone *zone);
3965 * Global mutex to protect against size modification of zonelists
3966 * as well as to serialize pageset setup for the new populated zone.
3968 DEFINE_MUTEX(zonelists_mutex);
3970 /* return values int ....just for stop_machine() */
3971 static int __build_all_zonelists(void *data)
3975 pg_data_t *self = data;
3978 memset(node_load, 0, sizeof(node_load));
3981 if (self && !node_online(self->node_id)) {
3982 build_zonelists(self);
3983 build_zonelist_cache(self);
3986 for_each_online_node(nid) {
3987 pg_data_t *pgdat = NODE_DATA(nid);
3989 build_zonelists(pgdat);
3990 build_zonelist_cache(pgdat);
3994 * Initialize the boot_pagesets that are going to be used
3995 * for bootstrapping processors. The real pagesets for
3996 * each zone will be allocated later when the per cpu
3997 * allocator is available.
3999 * boot_pagesets are used also for bootstrapping offline
4000 * cpus if the system is already booted because the pagesets
4001 * are needed to initialize allocators on a specific cpu too.
4002 * F.e. the percpu allocator needs the page allocator which
4003 * needs the percpu allocator in order to allocate its pagesets
4004 * (a chicken-egg dilemma).
4006 for_each_possible_cpu(cpu) {
4007 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4009 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4011 * We now know the "local memory node" for each node--
4012 * i.e., the node of the first zone in the generic zonelist.
4013 * Set up numa_mem percpu variable for on-line cpus. During
4014 * boot, only the boot cpu should be on-line; we'll init the
4015 * secondary cpus' numa_mem as they come on-line. During
4016 * node/memory hotplug, we'll fixup all on-line cpus.
4018 if (cpu_online(cpu))
4019 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4026 static noinline void __init
4027 build_all_zonelists_init(void)
4029 __build_all_zonelists(NULL);
4030 mminit_verify_zonelist();
4031 cpuset_init_current_mems_allowed();
4035 * Called with zonelists_mutex held always
4036 * unless system_state == SYSTEM_BOOTING.
4038 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4039 * [we're only called with non-NULL zone through __meminit paths] and
4040 * (2) call of __init annotated helper build_all_zonelists_init
4041 * [protected by SYSTEM_BOOTING].
4043 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4045 set_zonelist_order();
4047 if (system_state == SYSTEM_BOOTING) {
4048 build_all_zonelists_init();
4050 #ifdef CONFIG_MEMORY_HOTPLUG
4052 setup_zone_pageset(zone);
4054 /* we have to stop all cpus to guarantee there is no user
4056 stop_machine(__build_all_zonelists, pgdat, NULL);
4057 /* cpuset refresh routine should be here */
4059 vm_total_pages = nr_free_pagecache_pages();
4061 * Disable grouping by mobility if the number of pages in the
4062 * system is too low to allow the mechanism to work. It would be
4063 * more accurate, but expensive to check per-zone. This check is
4064 * made on memory-hotadd so a system can start with mobility
4065 * disabled and enable it later
4067 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4068 page_group_by_mobility_disabled = 1;
4070 page_group_by_mobility_disabled = 0;
4072 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4073 "Total pages: %ld\n",
4075 zonelist_order_name[current_zonelist_order],
4076 page_group_by_mobility_disabled ? "off" : "on",
4079 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4084 * Helper functions to size the waitqueue hash table.
4085 * Essentially these want to choose hash table sizes sufficiently
4086 * large so that collisions trying to wait on pages are rare.
4087 * But in fact, the number of active page waitqueues on typical
4088 * systems is ridiculously low, less than 200. So this is even
4089 * conservative, even though it seems large.
4091 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4092 * waitqueues, i.e. the size of the waitq table given the number of pages.
4094 #define PAGES_PER_WAITQUEUE 256
4096 #ifndef CONFIG_MEMORY_HOTPLUG
4097 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4099 unsigned long size = 1;
4101 pages /= PAGES_PER_WAITQUEUE;
4103 while (size < pages)
4107 * Once we have dozens or even hundreds of threads sleeping
4108 * on IO we've got bigger problems than wait queue collision.
4109 * Limit the size of the wait table to a reasonable size.
4111 size = min(size, 4096UL);
4113 return max(size, 4UL);
4117 * A zone's size might be changed by hot-add, so it is not possible to determine
4118 * a suitable size for its wait_table. So we use the maximum size now.
4120 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4122 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4123 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4124 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4126 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4127 * or more by the traditional way. (See above). It equals:
4129 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4130 * ia64(16K page size) : = ( 8G + 4M)byte.
4131 * powerpc (64K page size) : = (32G +16M)byte.
4133 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4140 * This is an integer logarithm so that shifts can be used later
4141 * to extract the more random high bits from the multiplicative
4142 * hash function before the remainder is taken.
4144 static inline unsigned long wait_table_bits(unsigned long size)
4150 * Check if a pageblock contains reserved pages
4152 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4156 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4157 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4164 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4165 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4166 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4167 * higher will lead to a bigger reserve which will get freed as contiguous
4168 * blocks as reclaim kicks in
4170 static void setup_zone_migrate_reserve(struct zone *zone)
4172 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4174 unsigned long block_migratetype;
4179 * Get the start pfn, end pfn and the number of blocks to reserve
4180 * We have to be careful to be aligned to pageblock_nr_pages to
4181 * make sure that we always check pfn_valid for the first page in
4184 start_pfn = zone->zone_start_pfn;
4185 end_pfn = zone_end_pfn(zone);
4186 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4187 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4191 * Reserve blocks are generally in place to help high-order atomic
4192 * allocations that are short-lived. A min_free_kbytes value that
4193 * would result in more than 2 reserve blocks for atomic allocations
4194 * is assumed to be in place to help anti-fragmentation for the
4195 * future allocation of hugepages at runtime.
4197 reserve = min(2, reserve);
4198 old_reserve = zone->nr_migrate_reserve_block;
4200 /* When memory hot-add, we almost always need to do nothing */
4201 if (reserve == old_reserve)
4203 zone->nr_migrate_reserve_block = reserve;
4205 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4206 if (!pfn_valid(pfn))
4208 page = pfn_to_page(pfn);
4210 /* Watch out for overlapping nodes */
4211 if (page_to_nid(page) != zone_to_nid(zone))
4214 block_migratetype = get_pageblock_migratetype(page);
4216 /* Only test what is necessary when the reserves are not met */
4219 * Blocks with reserved pages will never free, skip
4222 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4223 if (pageblock_is_reserved(pfn, block_end_pfn))
4226 /* If this block is reserved, account for it */
4227 if (block_migratetype == MIGRATE_RESERVE) {
4232 /* Suitable for reserving if this block is movable */
4233 if (block_migratetype == MIGRATE_MOVABLE) {
4234 set_pageblock_migratetype(page,
4236 move_freepages_block(zone, page,
4241 } else if (!old_reserve) {
4243 * At boot time we don't need to scan the whole zone
4244 * for turning off MIGRATE_RESERVE.
4250 * If the reserve is met and this is a previous reserved block,
4253 if (block_migratetype == MIGRATE_RESERVE) {
4254 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4255 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4261 * Initially all pages are reserved - free ones are freed
4262 * up by free_all_bootmem() once the early boot process is
4263 * done. Non-atomic initialization, single-pass.
4265 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4266 unsigned long start_pfn, enum memmap_context context)
4269 unsigned long end_pfn = start_pfn + size;
4273 if (highest_memmap_pfn < end_pfn - 1)
4274 highest_memmap_pfn = end_pfn - 1;
4276 z = &NODE_DATA(nid)->node_zones[zone];
4277 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4279 * There can be holes in boot-time mem_map[]s
4280 * handed to this function. They do not
4281 * exist on hotplugged memory.
4283 if (context == MEMMAP_EARLY) {
4284 if (!early_pfn_valid(pfn))
4286 if (!early_pfn_in_nid(pfn, nid))
4289 page = pfn_to_page(pfn);
4290 set_page_links(page, zone, nid, pfn);
4291 mminit_verify_page_links(page, zone, nid, pfn);
4292 init_page_count(page);
4293 page_mapcount_reset(page);
4294 page_cpupid_reset_last(page);
4295 SetPageReserved(page);
4297 * Mark the block movable so that blocks are reserved for
4298 * movable at startup. This will force kernel allocations
4299 * to reserve their blocks rather than leaking throughout
4300 * the address space during boot when many long-lived
4301 * kernel allocations are made. Later some blocks near
4302 * the start are marked MIGRATE_RESERVE by
4303 * setup_zone_migrate_reserve()
4305 * bitmap is created for zone's valid pfn range. but memmap
4306 * can be created for invalid pages (for alignment)
4307 * check here not to call set_pageblock_migratetype() against
4310 if ((z->zone_start_pfn <= pfn)
4311 && (pfn < zone_end_pfn(z))
4312 && !(pfn & (pageblock_nr_pages - 1)))
4313 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4315 INIT_LIST_HEAD(&page->lru);
4316 #ifdef WANT_PAGE_VIRTUAL
4317 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4318 if (!is_highmem_idx(zone))
4319 set_page_address(page, __va(pfn << PAGE_SHIFT));
4324 static void __meminit zone_init_free_lists(struct zone *zone)
4326 unsigned int order, t;
4327 for_each_migratetype_order(order, t) {
4328 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4329 zone->free_area[order].nr_free = 0;
4333 #ifndef __HAVE_ARCH_MEMMAP_INIT
4334 #define memmap_init(size, nid, zone, start_pfn) \
4335 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4338 static int zone_batchsize(struct zone *zone)
4344 * The per-cpu-pages pools are set to around 1000th of the
4345 * size of the zone. But no more than 1/2 of a meg.
4347 * OK, so we don't know how big the cache is. So guess.
4349 batch = zone->managed_pages / 1024;
4350 if (batch * PAGE_SIZE > 512 * 1024)
4351 batch = (512 * 1024) / PAGE_SIZE;
4352 batch /= 4; /* We effectively *= 4 below */
4357 * Clamp the batch to a 2^n - 1 value. Having a power
4358 * of 2 value was found to be more likely to have
4359 * suboptimal cache aliasing properties in some cases.
4361 * For example if 2 tasks are alternately allocating
4362 * batches of pages, one task can end up with a lot
4363 * of pages of one half of the possible page colors
4364 * and the other with pages of the other colors.
4366 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4371 /* The deferral and batching of frees should be suppressed under NOMMU
4374 * The problem is that NOMMU needs to be able to allocate large chunks
4375 * of contiguous memory as there's no hardware page translation to
4376 * assemble apparent contiguous memory from discontiguous pages.
4378 * Queueing large contiguous runs of pages for batching, however,
4379 * causes the pages to actually be freed in smaller chunks. As there
4380 * can be a significant delay between the individual batches being
4381 * recycled, this leads to the once large chunks of space being
4382 * fragmented and becoming unavailable for high-order allocations.
4389 * pcp->high and pcp->batch values are related and dependent on one another:
4390 * ->batch must never be higher then ->high.
4391 * The following function updates them in a safe manner without read side
4394 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4395 * those fields changing asynchronously (acording the the above rule).
4397 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4398 * outside of boot time (or some other assurance that no concurrent updaters
4401 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4402 unsigned long batch)
4404 /* start with a fail safe value for batch */
4408 /* Update high, then batch, in order */
4415 /* a companion to pageset_set_high() */
4416 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4418 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4421 static void pageset_init(struct per_cpu_pageset *p)
4423 struct per_cpu_pages *pcp;
4426 memset(p, 0, sizeof(*p));
4430 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4431 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4434 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4437 pageset_set_batch(p, batch);
4441 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4442 * to the value high for the pageset p.
4444 static void pageset_set_high(struct per_cpu_pageset *p,
4447 unsigned long batch = max(1UL, high / 4);
4448 if ((high / 4) > (PAGE_SHIFT * 8))
4449 batch = PAGE_SHIFT * 8;
4451 pageset_update(&p->pcp, high, batch);
4454 static void pageset_set_high_and_batch(struct zone *zone,
4455 struct per_cpu_pageset *pcp)
4457 if (percpu_pagelist_fraction)
4458 pageset_set_high(pcp,
4459 (zone->managed_pages /
4460 percpu_pagelist_fraction));
4462 pageset_set_batch(pcp, zone_batchsize(zone));
4465 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4467 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4470 pageset_set_high_and_batch(zone, pcp);
4473 static void __meminit setup_zone_pageset(struct zone *zone)
4476 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4477 for_each_possible_cpu(cpu)
4478 zone_pageset_init(zone, cpu);
4482 * Allocate per cpu pagesets and initialize them.
4483 * Before this call only boot pagesets were available.
4485 void __init setup_per_cpu_pageset(void)
4489 for_each_populated_zone(zone)
4490 setup_zone_pageset(zone);
4493 static noinline __init_refok
4494 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4500 * The per-page waitqueue mechanism uses hashed waitqueues
4503 zone->wait_table_hash_nr_entries =
4504 wait_table_hash_nr_entries(zone_size_pages);
4505 zone->wait_table_bits =
4506 wait_table_bits(zone->wait_table_hash_nr_entries);
4507 alloc_size = zone->wait_table_hash_nr_entries
4508 * sizeof(wait_queue_head_t);
4510 if (!slab_is_available()) {
4511 zone->wait_table = (wait_queue_head_t *)
4512 memblock_virt_alloc_node_nopanic(
4513 alloc_size, zone->zone_pgdat->node_id);
4516 * This case means that a zone whose size was 0 gets new memory
4517 * via memory hot-add.
4518 * But it may be the case that a new node was hot-added. In
4519 * this case vmalloc() will not be able to use this new node's
4520 * memory - this wait_table must be initialized to use this new
4521 * node itself as well.
4522 * To use this new node's memory, further consideration will be
4525 zone->wait_table = vmalloc(alloc_size);
4527 if (!zone->wait_table)
4530 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4531 init_waitqueue_head(zone->wait_table + i);
4536 static __meminit void zone_pcp_init(struct zone *zone)
4539 * per cpu subsystem is not up at this point. The following code
4540 * relies on the ability of the linker to provide the
4541 * offset of a (static) per cpu variable into the per cpu area.
4543 zone->pageset = &boot_pageset;
4545 if (populated_zone(zone))
4546 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4547 zone->name, zone->present_pages,
4548 zone_batchsize(zone));
4551 int __meminit init_currently_empty_zone(struct zone *zone,
4552 unsigned long zone_start_pfn,
4554 enum memmap_context context)
4556 struct pglist_data *pgdat = zone->zone_pgdat;
4558 ret = zone_wait_table_init(zone, size);
4561 pgdat->nr_zones = zone_idx(zone) + 1;
4563 zone->zone_start_pfn = zone_start_pfn;
4565 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4566 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4568 (unsigned long)zone_idx(zone),
4569 zone_start_pfn, (zone_start_pfn + size));
4571 zone_init_free_lists(zone);
4576 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4577 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4579 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4581 int __meminit __early_pfn_to_nid(unsigned long pfn)
4583 unsigned long start_pfn, end_pfn;
4586 * NOTE: The following SMP-unsafe globals are only used early in boot
4587 * when the kernel is running single-threaded.
4589 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4590 static int __meminitdata last_nid;
4592 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4595 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4597 last_start_pfn = start_pfn;
4598 last_end_pfn = end_pfn;
4604 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4606 int __meminit early_pfn_to_nid(unsigned long pfn)
4610 nid = __early_pfn_to_nid(pfn);
4613 /* just returns 0 */
4617 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4618 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4622 nid = __early_pfn_to_nid(pfn);
4623 if (nid >= 0 && nid != node)
4630 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4631 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4632 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4634 * If an architecture guarantees that all ranges registered contain no holes
4635 * and may be freed, this this function may be used instead of calling
4636 * memblock_free_early_nid() manually.
4638 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4640 unsigned long start_pfn, end_pfn;
4643 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4644 start_pfn = min(start_pfn, max_low_pfn);
4645 end_pfn = min(end_pfn, max_low_pfn);
4647 if (start_pfn < end_pfn)
4648 memblock_free_early_nid(PFN_PHYS(start_pfn),
4649 (end_pfn - start_pfn) << PAGE_SHIFT,
4655 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4656 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4658 * If an architecture guarantees that all ranges registered contain no holes and may
4659 * be freed, this function may be used instead of calling memory_present() manually.
4661 void __init sparse_memory_present_with_active_regions(int nid)
4663 unsigned long start_pfn, end_pfn;
4666 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4667 memory_present(this_nid, start_pfn, end_pfn);
4671 * get_pfn_range_for_nid - Return the start and end page frames for a node
4672 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4673 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4674 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4676 * It returns the start and end page frame of a node based on information
4677 * provided by memblock_set_node(). If called for a node
4678 * with no available memory, a warning is printed and the start and end
4681 void __meminit get_pfn_range_for_nid(unsigned int nid,
4682 unsigned long *start_pfn, unsigned long *end_pfn)
4684 unsigned long this_start_pfn, this_end_pfn;
4690 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4691 *start_pfn = min(*start_pfn, this_start_pfn);
4692 *end_pfn = max(*end_pfn, this_end_pfn);
4695 if (*start_pfn == -1UL)
4700 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4701 * assumption is made that zones within a node are ordered in monotonic
4702 * increasing memory addresses so that the "highest" populated zone is used
4704 static void __init find_usable_zone_for_movable(void)
4707 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4708 if (zone_index == ZONE_MOVABLE)
4711 if (arch_zone_highest_possible_pfn[zone_index] >
4712 arch_zone_lowest_possible_pfn[zone_index])
4716 VM_BUG_ON(zone_index == -1);
4717 movable_zone = zone_index;
4721 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4722 * because it is sized independent of architecture. Unlike the other zones,
4723 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4724 * in each node depending on the size of each node and how evenly kernelcore
4725 * is distributed. This helper function adjusts the zone ranges
4726 * provided by the architecture for a given node by using the end of the
4727 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4728 * zones within a node are in order of monotonic increases memory addresses
4730 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4731 unsigned long zone_type,
4732 unsigned long node_start_pfn,
4733 unsigned long node_end_pfn,
4734 unsigned long *zone_start_pfn,
4735 unsigned long *zone_end_pfn)
4737 /* Only adjust if ZONE_MOVABLE is on this node */
4738 if (zone_movable_pfn[nid]) {
4739 /* Size ZONE_MOVABLE */
4740 if (zone_type == ZONE_MOVABLE) {
4741 *zone_start_pfn = zone_movable_pfn[nid];
4742 *zone_end_pfn = min(node_end_pfn,
4743 arch_zone_highest_possible_pfn[movable_zone]);
4745 /* Adjust for ZONE_MOVABLE starting within this range */
4746 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4747 *zone_end_pfn > zone_movable_pfn[nid]) {
4748 *zone_end_pfn = zone_movable_pfn[nid];
4750 /* Check if this whole range is within ZONE_MOVABLE */
4751 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4752 *zone_start_pfn = *zone_end_pfn;
4757 * Return the number of pages a zone spans in a node, including holes
4758 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4760 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4761 unsigned long zone_type,
4762 unsigned long node_start_pfn,
4763 unsigned long node_end_pfn,
4764 unsigned long *ignored)
4766 unsigned long zone_start_pfn, zone_end_pfn;
4768 /* Get the start and end of the zone */
4769 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4770 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4771 adjust_zone_range_for_zone_movable(nid, zone_type,
4772 node_start_pfn, node_end_pfn,
4773 &zone_start_pfn, &zone_end_pfn);
4775 /* Check that this node has pages within the zone's required range */
4776 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4779 /* Move the zone boundaries inside the node if necessary */
4780 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4781 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4783 /* Return the spanned pages */
4784 return zone_end_pfn - zone_start_pfn;
4788 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4789 * then all holes in the requested range will be accounted for.
4791 unsigned long __meminit __absent_pages_in_range(int nid,
4792 unsigned long range_start_pfn,
4793 unsigned long range_end_pfn)
4795 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4796 unsigned long start_pfn, end_pfn;
4799 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4800 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4801 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4802 nr_absent -= end_pfn - start_pfn;
4808 * absent_pages_in_range - Return number of page frames in holes within a range
4809 * @start_pfn: The start PFN to start searching for holes
4810 * @end_pfn: The end PFN to stop searching for holes
4812 * It returns the number of pages frames in memory holes within a range.
4814 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4815 unsigned long end_pfn)
4817 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4820 /* Return the number of page frames in holes in a zone on a node */
4821 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4822 unsigned long zone_type,
4823 unsigned long node_start_pfn,
4824 unsigned long node_end_pfn,
4825 unsigned long *ignored)
4827 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4828 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4829 unsigned long zone_start_pfn, zone_end_pfn;
4831 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4832 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4834 adjust_zone_range_for_zone_movable(nid, zone_type,
4835 node_start_pfn, node_end_pfn,
4836 &zone_start_pfn, &zone_end_pfn);
4837 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4840 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4841 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4842 unsigned long zone_type,
4843 unsigned long node_start_pfn,
4844 unsigned long node_end_pfn,
4845 unsigned long *zones_size)
4847 return zones_size[zone_type];
4850 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4851 unsigned long zone_type,
4852 unsigned long node_start_pfn,
4853 unsigned long node_end_pfn,
4854 unsigned long *zholes_size)
4859 return zholes_size[zone_type];
4862 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4864 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4865 unsigned long node_start_pfn,
4866 unsigned long node_end_pfn,
4867 unsigned long *zones_size,
4868 unsigned long *zholes_size)
4870 unsigned long realtotalpages, totalpages = 0;
4873 for (i = 0; i < MAX_NR_ZONES; i++)
4874 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4878 pgdat->node_spanned_pages = totalpages;
4880 realtotalpages = totalpages;
4881 for (i = 0; i < MAX_NR_ZONES; i++)
4883 zone_absent_pages_in_node(pgdat->node_id, i,
4884 node_start_pfn, node_end_pfn,
4886 pgdat->node_present_pages = realtotalpages;
4887 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4891 #ifndef CONFIG_SPARSEMEM
4893 * Calculate the size of the zone->blockflags rounded to an unsigned long
4894 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4895 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4896 * round what is now in bits to nearest long in bits, then return it in
4899 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4901 unsigned long usemapsize;
4903 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4904 usemapsize = roundup(zonesize, pageblock_nr_pages);
4905 usemapsize = usemapsize >> pageblock_order;
4906 usemapsize *= NR_PAGEBLOCK_BITS;
4907 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4909 return usemapsize / 8;
4912 static void __init setup_usemap(struct pglist_data *pgdat,
4914 unsigned long zone_start_pfn,
4915 unsigned long zonesize)
4917 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4918 zone->pageblock_flags = NULL;
4920 zone->pageblock_flags =
4921 memblock_virt_alloc_node_nopanic(usemapsize,
4925 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4926 unsigned long zone_start_pfn, unsigned long zonesize) {}
4927 #endif /* CONFIG_SPARSEMEM */
4929 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4931 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4932 void __paginginit set_pageblock_order(void)
4936 /* Check that pageblock_nr_pages has not already been setup */
4937 if (pageblock_order)
4940 if (HPAGE_SHIFT > PAGE_SHIFT)
4941 order = HUGETLB_PAGE_ORDER;
4943 order = MAX_ORDER - 1;
4946 * Assume the largest contiguous order of interest is a huge page.
4947 * This value may be variable depending on boot parameters on IA64 and
4950 pageblock_order = order;
4952 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4955 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4956 * is unused as pageblock_order is set at compile-time. See
4957 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4960 void __paginginit set_pageblock_order(void)
4964 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4966 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4967 unsigned long present_pages)
4969 unsigned long pages = spanned_pages;
4972 * Provide a more accurate estimation if there are holes within
4973 * the zone and SPARSEMEM is in use. If there are holes within the
4974 * zone, each populated memory region may cost us one or two extra
4975 * memmap pages due to alignment because memmap pages for each
4976 * populated regions may not naturally algined on page boundary.
4977 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4979 if (spanned_pages > present_pages + (present_pages >> 4) &&
4980 IS_ENABLED(CONFIG_SPARSEMEM))
4981 pages = present_pages;
4983 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4987 * Set up the zone data structures:
4988 * - mark all pages reserved
4989 * - mark all memory queues empty
4990 * - clear the memory bitmaps
4992 * NOTE: pgdat should get zeroed by caller.
4994 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4995 unsigned long node_start_pfn, unsigned long node_end_pfn,
4996 unsigned long *zones_size, unsigned long *zholes_size)
4999 int nid = pgdat->node_id;
5000 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5003 pgdat_resize_init(pgdat);
5004 #ifdef CONFIG_NUMA_BALANCING
5005 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5006 pgdat->numabalancing_migrate_nr_pages = 0;
5007 pgdat->numabalancing_migrate_next_window = jiffies;
5009 init_waitqueue_head(&pgdat->kswapd_wait);
5010 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5011 pgdat_page_ext_init(pgdat);
5013 for (j = 0; j < MAX_NR_ZONES; j++) {
5014 struct zone *zone = pgdat->node_zones + j;
5015 unsigned long size, realsize, freesize, memmap_pages;
5017 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
5018 node_end_pfn, zones_size);
5019 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
5025 * Adjust freesize so that it accounts for how much memory
5026 * is used by this zone for memmap. This affects the watermark
5027 * and per-cpu initialisations
5029 memmap_pages = calc_memmap_size(size, realsize);
5030 if (!is_highmem_idx(j)) {
5031 if (freesize >= memmap_pages) {
5032 freesize -= memmap_pages;
5035 " %s zone: %lu pages used for memmap\n",
5036 zone_names[j], memmap_pages);
5039 " %s zone: %lu pages exceeds freesize %lu\n",
5040 zone_names[j], memmap_pages, freesize);
5043 /* Account for reserved pages */
5044 if (j == 0 && freesize > dma_reserve) {
5045 freesize -= dma_reserve;
5046 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5047 zone_names[0], dma_reserve);
5050 if (!is_highmem_idx(j))
5051 nr_kernel_pages += freesize;
5052 /* Charge for highmem memmap if there are enough kernel pages */
5053 else if (nr_kernel_pages > memmap_pages * 2)
5054 nr_kernel_pages -= memmap_pages;
5055 nr_all_pages += freesize;
5057 zone->spanned_pages = size;
5058 zone->present_pages = realsize;
5060 * Set an approximate value for lowmem here, it will be adjusted
5061 * when the bootmem allocator frees pages into the buddy system.
5062 * And all highmem pages will be managed by the buddy system.
5064 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5067 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5069 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5071 zone->name = zone_names[j];
5072 spin_lock_init(&zone->lock);
5073 spin_lock_init(&zone->lru_lock);
5074 zone_seqlock_init(zone);
5075 zone->zone_pgdat = pgdat;
5076 zone_pcp_init(zone);
5078 /* For bootup, initialized properly in watermark setup */
5079 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5081 lruvec_init(&zone->lruvec);
5085 set_pageblock_order();
5086 setup_usemap(pgdat, zone, zone_start_pfn, size);
5087 ret = init_currently_empty_zone(zone, zone_start_pfn,
5088 size, MEMMAP_EARLY);
5090 memmap_init(size, nid, j, zone_start_pfn);
5091 zone_start_pfn += size;
5095 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5097 /* Skip empty nodes */
5098 if (!pgdat->node_spanned_pages)
5101 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5102 /* ia64 gets its own node_mem_map, before this, without bootmem */
5103 if (!pgdat->node_mem_map) {
5104 unsigned long size, start, end;
5108 * The zone's endpoints aren't required to be MAX_ORDER
5109 * aligned but the node_mem_map endpoints must be in order
5110 * for the buddy allocator to function correctly.
5112 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5113 end = pgdat_end_pfn(pgdat);
5114 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5115 size = (end - start) * sizeof(struct page);
5116 map = alloc_remap(pgdat->node_id, size);
5118 map = memblock_virt_alloc_node_nopanic(size,
5120 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5122 #ifndef CONFIG_NEED_MULTIPLE_NODES
5124 * With no DISCONTIG, the global mem_map is just set as node 0's
5126 if (pgdat == NODE_DATA(0)) {
5127 mem_map = NODE_DATA(0)->node_mem_map;
5128 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5129 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5130 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5131 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5134 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5137 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5138 unsigned long node_start_pfn, unsigned long *zholes_size)
5140 pg_data_t *pgdat = NODE_DATA(nid);
5141 unsigned long start_pfn = 0;
5142 unsigned long end_pfn = 0;
5144 /* pg_data_t should be reset to zero when it's allocated */
5145 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5147 pgdat->node_id = nid;
5148 pgdat->node_start_pfn = node_start_pfn;
5149 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5150 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5151 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5152 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5154 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5155 zones_size, zholes_size);
5157 alloc_node_mem_map(pgdat);
5158 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5159 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5160 nid, (unsigned long)pgdat,
5161 (unsigned long)pgdat->node_mem_map);
5164 free_area_init_core(pgdat, start_pfn, end_pfn,
5165 zones_size, zholes_size);
5168 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5170 #if MAX_NUMNODES > 1
5172 * Figure out the number of possible node ids.
5174 void __init setup_nr_node_ids(void)
5177 unsigned int highest = 0;
5179 for_each_node_mask(node, node_possible_map)
5181 nr_node_ids = highest + 1;
5186 * node_map_pfn_alignment - determine the maximum internode alignment
5188 * This function should be called after node map is populated and sorted.
5189 * It calculates the maximum power of two alignment which can distinguish
5192 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5193 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5194 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5195 * shifted, 1GiB is enough and this function will indicate so.
5197 * This is used to test whether pfn -> nid mapping of the chosen memory
5198 * model has fine enough granularity to avoid incorrect mapping for the
5199 * populated node map.
5201 * Returns the determined alignment in pfn's. 0 if there is no alignment
5202 * requirement (single node).
5204 unsigned long __init node_map_pfn_alignment(void)
5206 unsigned long accl_mask = 0, last_end = 0;
5207 unsigned long start, end, mask;
5211 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5212 if (!start || last_nid < 0 || last_nid == nid) {
5219 * Start with a mask granular enough to pin-point to the
5220 * start pfn and tick off bits one-by-one until it becomes
5221 * too coarse to separate the current node from the last.
5223 mask = ~((1 << __ffs(start)) - 1);
5224 while (mask && last_end <= (start & (mask << 1)))
5227 /* accumulate all internode masks */
5231 /* convert mask to number of pages */
5232 return ~accl_mask + 1;
5235 /* Find the lowest pfn for a node */
5236 static unsigned long __init find_min_pfn_for_node(int nid)
5238 unsigned long min_pfn = ULONG_MAX;
5239 unsigned long start_pfn;
5242 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5243 min_pfn = min(min_pfn, start_pfn);
5245 if (min_pfn == ULONG_MAX) {
5247 "Could not find start_pfn for node %d\n", nid);
5255 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5257 * It returns the minimum PFN based on information provided via
5258 * memblock_set_node().
5260 unsigned long __init find_min_pfn_with_active_regions(void)
5262 return find_min_pfn_for_node(MAX_NUMNODES);
5266 * early_calculate_totalpages()
5267 * Sum pages in active regions for movable zone.
5268 * Populate N_MEMORY for calculating usable_nodes.
5270 static unsigned long __init early_calculate_totalpages(void)
5272 unsigned long totalpages = 0;
5273 unsigned long start_pfn, end_pfn;
5276 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5277 unsigned long pages = end_pfn - start_pfn;
5279 totalpages += pages;
5281 node_set_state(nid, N_MEMORY);
5287 * Find the PFN the Movable zone begins in each node. Kernel memory
5288 * is spread evenly between nodes as long as the nodes have enough
5289 * memory. When they don't, some nodes will have more kernelcore than
5292 static void __init find_zone_movable_pfns_for_nodes(void)
5295 unsigned long usable_startpfn;
5296 unsigned long kernelcore_node, kernelcore_remaining;
5297 /* save the state before borrow the nodemask */
5298 nodemask_t saved_node_state = node_states[N_MEMORY];
5299 unsigned long totalpages = early_calculate_totalpages();
5300 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5301 struct memblock_region *r;
5303 /* Need to find movable_zone earlier when movable_node is specified. */
5304 find_usable_zone_for_movable();
5307 * If movable_node is specified, ignore kernelcore and movablecore
5310 if (movable_node_is_enabled()) {
5311 for_each_memblock(memory, r) {
5312 if (!memblock_is_hotpluggable(r))
5317 usable_startpfn = PFN_DOWN(r->base);
5318 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5319 min(usable_startpfn, zone_movable_pfn[nid]) :
5327 * If movablecore=nn[KMG] was specified, calculate what size of
5328 * kernelcore that corresponds so that memory usable for
5329 * any allocation type is evenly spread. If both kernelcore
5330 * and movablecore are specified, then the value of kernelcore
5331 * will be used for required_kernelcore if it's greater than
5332 * what movablecore would have allowed.
5334 if (required_movablecore) {
5335 unsigned long corepages;
5338 * Round-up so that ZONE_MOVABLE is at least as large as what
5339 * was requested by the user
5341 required_movablecore =
5342 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5343 corepages = totalpages - required_movablecore;
5345 required_kernelcore = max(required_kernelcore, corepages);
5348 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5349 if (!required_kernelcore)
5352 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5353 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5356 /* Spread kernelcore memory as evenly as possible throughout nodes */
5357 kernelcore_node = required_kernelcore / usable_nodes;
5358 for_each_node_state(nid, N_MEMORY) {
5359 unsigned long start_pfn, end_pfn;
5362 * Recalculate kernelcore_node if the division per node
5363 * now exceeds what is necessary to satisfy the requested
5364 * amount of memory for the kernel
5366 if (required_kernelcore < kernelcore_node)
5367 kernelcore_node = required_kernelcore / usable_nodes;
5370 * As the map is walked, we track how much memory is usable
5371 * by the kernel using kernelcore_remaining. When it is
5372 * 0, the rest of the node is usable by ZONE_MOVABLE
5374 kernelcore_remaining = kernelcore_node;
5376 /* Go through each range of PFNs within this node */
5377 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5378 unsigned long size_pages;
5380 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5381 if (start_pfn >= end_pfn)
5384 /* Account for what is only usable for kernelcore */
5385 if (start_pfn < usable_startpfn) {
5386 unsigned long kernel_pages;
5387 kernel_pages = min(end_pfn, usable_startpfn)
5390 kernelcore_remaining -= min(kernel_pages,
5391 kernelcore_remaining);
5392 required_kernelcore -= min(kernel_pages,
5393 required_kernelcore);
5395 /* Continue if range is now fully accounted */
5396 if (end_pfn <= usable_startpfn) {
5399 * Push zone_movable_pfn to the end so
5400 * that if we have to rebalance
5401 * kernelcore across nodes, we will
5402 * not double account here
5404 zone_movable_pfn[nid] = end_pfn;
5407 start_pfn = usable_startpfn;
5411 * The usable PFN range for ZONE_MOVABLE is from
5412 * start_pfn->end_pfn. Calculate size_pages as the
5413 * number of pages used as kernelcore
5415 size_pages = end_pfn - start_pfn;
5416 if (size_pages > kernelcore_remaining)
5417 size_pages = kernelcore_remaining;
5418 zone_movable_pfn[nid] = start_pfn + size_pages;
5421 * Some kernelcore has been met, update counts and
5422 * break if the kernelcore for this node has been
5425 required_kernelcore -= min(required_kernelcore,
5427 kernelcore_remaining -= size_pages;
5428 if (!kernelcore_remaining)
5434 * If there is still required_kernelcore, we do another pass with one
5435 * less node in the count. This will push zone_movable_pfn[nid] further
5436 * along on the nodes that still have memory until kernelcore is
5440 if (usable_nodes && required_kernelcore > usable_nodes)
5444 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5445 for (nid = 0; nid < MAX_NUMNODES; nid++)
5446 zone_movable_pfn[nid] =
5447 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5450 /* restore the node_state */
5451 node_states[N_MEMORY] = saved_node_state;
5454 /* Any regular or high memory on that node ? */
5455 static void check_for_memory(pg_data_t *pgdat, int nid)
5457 enum zone_type zone_type;
5459 if (N_MEMORY == N_NORMAL_MEMORY)
5462 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5463 struct zone *zone = &pgdat->node_zones[zone_type];
5464 if (populated_zone(zone)) {
5465 node_set_state(nid, N_HIGH_MEMORY);
5466 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5467 zone_type <= ZONE_NORMAL)
5468 node_set_state(nid, N_NORMAL_MEMORY);
5475 * free_area_init_nodes - Initialise all pg_data_t and zone data
5476 * @max_zone_pfn: an array of max PFNs for each zone
5478 * This will call free_area_init_node() for each active node in the system.
5479 * Using the page ranges provided by memblock_set_node(), the size of each
5480 * zone in each node and their holes is calculated. If the maximum PFN
5481 * between two adjacent zones match, it is assumed that the zone is empty.
5482 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5483 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5484 * starts where the previous one ended. For example, ZONE_DMA32 starts
5485 * at arch_max_dma_pfn.
5487 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5489 unsigned long start_pfn, end_pfn;
5492 /* Record where the zone boundaries are */
5493 memset(arch_zone_lowest_possible_pfn, 0,
5494 sizeof(arch_zone_lowest_possible_pfn));
5495 memset(arch_zone_highest_possible_pfn, 0,
5496 sizeof(arch_zone_highest_possible_pfn));
5497 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5498 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5499 for (i = 1; i < MAX_NR_ZONES; i++) {
5500 if (i == ZONE_MOVABLE)
5502 arch_zone_lowest_possible_pfn[i] =
5503 arch_zone_highest_possible_pfn[i-1];
5504 arch_zone_highest_possible_pfn[i] =
5505 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5507 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5508 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5510 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5511 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5512 find_zone_movable_pfns_for_nodes();
5514 /* Print out the zone ranges */
5515 pr_info("Zone ranges:\n");
5516 for (i = 0; i < MAX_NR_ZONES; i++) {
5517 if (i == ZONE_MOVABLE)
5519 pr_info(" %-8s ", zone_names[i]);
5520 if (arch_zone_lowest_possible_pfn[i] ==
5521 arch_zone_highest_possible_pfn[i])
5524 pr_cont("[mem %#018Lx-%#018Lx]\n",
5525 (u64)arch_zone_lowest_possible_pfn[i]
5527 ((u64)arch_zone_highest_possible_pfn[i]
5528 << PAGE_SHIFT) - 1);
5531 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5532 pr_info("Movable zone start for each node\n");
5533 for (i = 0; i < MAX_NUMNODES; i++) {
5534 if (zone_movable_pfn[i])
5535 pr_info(" Node %d: %#018Lx\n", i,
5536 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5539 /* Print out the early node map */
5540 pr_info("Early memory node ranges\n");
5541 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5542 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5543 (u64)start_pfn << PAGE_SHIFT,
5544 ((u64)end_pfn << PAGE_SHIFT) - 1);
5546 /* Initialise every node */
5547 mminit_verify_pageflags_layout();
5548 setup_nr_node_ids();
5549 for_each_online_node(nid) {
5550 pg_data_t *pgdat = NODE_DATA(nid);
5551 free_area_init_node(nid, NULL,
5552 find_min_pfn_for_node(nid), NULL);
5554 /* Any memory on that node */
5555 if (pgdat->node_present_pages)
5556 node_set_state(nid, N_MEMORY);
5557 check_for_memory(pgdat, nid);
5561 static int __init cmdline_parse_core(char *p, unsigned long *core)
5563 unsigned long long coremem;
5567 coremem = memparse(p, &p);
5568 *core = coremem >> PAGE_SHIFT;
5570 /* Paranoid check that UL is enough for the coremem value */
5571 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5577 * kernelcore=size sets the amount of memory for use for allocations that
5578 * cannot be reclaimed or migrated.
5580 static int __init cmdline_parse_kernelcore(char *p)
5582 return cmdline_parse_core(p, &required_kernelcore);
5586 * movablecore=size sets the amount of memory for use for allocations that
5587 * can be reclaimed or migrated.
5589 static int __init cmdline_parse_movablecore(char *p)
5591 return cmdline_parse_core(p, &required_movablecore);
5594 early_param("kernelcore", cmdline_parse_kernelcore);
5595 early_param("movablecore", cmdline_parse_movablecore);
5597 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5599 void adjust_managed_page_count(struct page *page, long count)
5601 spin_lock(&managed_page_count_lock);
5602 page_zone(page)->managed_pages += count;
5603 totalram_pages += count;
5604 #ifdef CONFIG_HIGHMEM
5605 if (PageHighMem(page))
5606 totalhigh_pages += count;
5608 spin_unlock(&managed_page_count_lock);
5610 EXPORT_SYMBOL(adjust_managed_page_count);
5612 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5615 unsigned long pages = 0;
5617 start = (void *)PAGE_ALIGN((unsigned long)start);
5618 end = (void *)((unsigned long)end & PAGE_MASK);
5619 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5620 if ((unsigned int)poison <= 0xFF)
5621 memset(pos, poison, PAGE_SIZE);
5622 free_reserved_page(virt_to_page(pos));
5626 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5627 s, pages << (PAGE_SHIFT - 10), start, end);
5631 EXPORT_SYMBOL(free_reserved_area);
5633 #ifdef CONFIG_HIGHMEM
5634 void free_highmem_page(struct page *page)
5636 __free_reserved_page(page);
5638 page_zone(page)->managed_pages++;
5644 void __init mem_init_print_info(const char *str)
5646 unsigned long physpages, codesize, datasize, rosize, bss_size;
5647 unsigned long init_code_size, init_data_size;
5649 physpages = get_num_physpages();
5650 codesize = _etext - _stext;
5651 datasize = _edata - _sdata;
5652 rosize = __end_rodata - __start_rodata;
5653 bss_size = __bss_stop - __bss_start;
5654 init_data_size = __init_end - __init_begin;
5655 init_code_size = _einittext - _sinittext;
5658 * Detect special cases and adjust section sizes accordingly:
5659 * 1) .init.* may be embedded into .data sections
5660 * 2) .init.text.* may be out of [__init_begin, __init_end],
5661 * please refer to arch/tile/kernel/vmlinux.lds.S.
5662 * 3) .rodata.* may be embedded into .text or .data sections.
5664 #define adj_init_size(start, end, size, pos, adj) \
5666 if (start <= pos && pos < end && size > adj) \
5670 adj_init_size(__init_begin, __init_end, init_data_size,
5671 _sinittext, init_code_size);
5672 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5673 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5674 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5675 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5677 #undef adj_init_size
5679 pr_info("Memory: %luK/%luK available "
5680 "(%luK kernel code, %luK rwdata, %luK rodata, "
5681 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5682 #ifdef CONFIG_HIGHMEM
5686 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5687 codesize >> 10, datasize >> 10, rosize >> 10,
5688 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5689 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5690 totalcma_pages << (PAGE_SHIFT-10),
5691 #ifdef CONFIG_HIGHMEM
5692 totalhigh_pages << (PAGE_SHIFT-10),
5694 str ? ", " : "", str ? str : "");
5698 * set_dma_reserve - set the specified number of pages reserved in the first zone
5699 * @new_dma_reserve: The number of pages to mark reserved
5701 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5702 * In the DMA zone, a significant percentage may be consumed by kernel image
5703 * and other unfreeable allocations which can skew the watermarks badly. This
5704 * function may optionally be used to account for unfreeable pages in the
5705 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5706 * smaller per-cpu batchsize.
5708 void __init set_dma_reserve(unsigned long new_dma_reserve)
5710 dma_reserve = new_dma_reserve;
5713 void __init free_area_init(unsigned long *zones_size)
5715 free_area_init_node(0, zones_size,
5716 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5719 static int page_alloc_cpu_notify(struct notifier_block *self,
5720 unsigned long action, void *hcpu)
5722 int cpu = (unsigned long)hcpu;
5724 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5725 lru_add_drain_cpu(cpu);
5729 * Spill the event counters of the dead processor
5730 * into the current processors event counters.
5731 * This artificially elevates the count of the current
5734 vm_events_fold_cpu(cpu);
5737 * Zero the differential counters of the dead processor
5738 * so that the vm statistics are consistent.
5740 * This is only okay since the processor is dead and cannot
5741 * race with what we are doing.
5743 cpu_vm_stats_fold(cpu);
5748 void __init page_alloc_init(void)
5750 hotcpu_notifier(page_alloc_cpu_notify, 0);
5754 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5755 * or min_free_kbytes changes.
5757 static void calculate_totalreserve_pages(void)
5759 struct pglist_data *pgdat;
5760 unsigned long reserve_pages = 0;
5761 enum zone_type i, j;
5763 for_each_online_pgdat(pgdat) {
5764 for (i = 0; i < MAX_NR_ZONES; i++) {
5765 struct zone *zone = pgdat->node_zones + i;
5768 /* Find valid and maximum lowmem_reserve in the zone */
5769 for (j = i; j < MAX_NR_ZONES; j++) {
5770 if (zone->lowmem_reserve[j] > max)
5771 max = zone->lowmem_reserve[j];
5774 /* we treat the high watermark as reserved pages. */
5775 max += high_wmark_pages(zone);
5777 if (max > zone->managed_pages)
5778 max = zone->managed_pages;
5779 reserve_pages += max;
5781 * Lowmem reserves are not available to
5782 * GFP_HIGHUSER page cache allocations and
5783 * kswapd tries to balance zones to their high
5784 * watermark. As a result, neither should be
5785 * regarded as dirtyable memory, to prevent a
5786 * situation where reclaim has to clean pages
5787 * in order to balance the zones.
5789 zone->dirty_balance_reserve = max;
5792 dirty_balance_reserve = reserve_pages;
5793 totalreserve_pages = reserve_pages;
5797 * setup_per_zone_lowmem_reserve - called whenever
5798 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5799 * has a correct pages reserved value, so an adequate number of
5800 * pages are left in the zone after a successful __alloc_pages().
5802 static void setup_per_zone_lowmem_reserve(void)
5804 struct pglist_data *pgdat;
5805 enum zone_type j, idx;
5807 for_each_online_pgdat(pgdat) {
5808 for (j = 0; j < MAX_NR_ZONES; j++) {
5809 struct zone *zone = pgdat->node_zones + j;
5810 unsigned long managed_pages = zone->managed_pages;
5812 zone->lowmem_reserve[j] = 0;
5816 struct zone *lower_zone;
5820 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5821 sysctl_lowmem_reserve_ratio[idx] = 1;
5823 lower_zone = pgdat->node_zones + idx;
5824 lower_zone->lowmem_reserve[j] = managed_pages /
5825 sysctl_lowmem_reserve_ratio[idx];
5826 managed_pages += lower_zone->managed_pages;
5831 /* update totalreserve_pages */
5832 calculate_totalreserve_pages();
5835 static void __setup_per_zone_wmarks(void)
5837 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5838 unsigned long lowmem_pages = 0;
5840 unsigned long flags;
5842 /* Calculate total number of !ZONE_HIGHMEM pages */
5843 for_each_zone(zone) {
5844 if (!is_highmem(zone))
5845 lowmem_pages += zone->managed_pages;
5848 for_each_zone(zone) {
5851 spin_lock_irqsave(&zone->lock, flags);
5852 tmp = (u64)pages_min * zone->managed_pages;
5853 do_div(tmp, lowmem_pages);
5854 if (is_highmem(zone)) {
5856 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5857 * need highmem pages, so cap pages_min to a small
5860 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5861 * deltas control asynch page reclaim, and so should
5862 * not be capped for highmem.
5864 unsigned long min_pages;
5866 min_pages = zone->managed_pages / 1024;
5867 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5868 zone->watermark[WMARK_MIN] = min_pages;
5871 * If it's a lowmem zone, reserve a number of pages
5872 * proportionate to the zone's size.
5874 zone->watermark[WMARK_MIN] = tmp;
5877 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5878 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5880 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5881 high_wmark_pages(zone) - low_wmark_pages(zone) -
5882 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5884 setup_zone_migrate_reserve(zone);
5885 spin_unlock_irqrestore(&zone->lock, flags);
5888 /* update totalreserve_pages */
5889 calculate_totalreserve_pages();
5893 * setup_per_zone_wmarks - called when min_free_kbytes changes
5894 * or when memory is hot-{added|removed}
5896 * Ensures that the watermark[min,low,high] values for each zone are set
5897 * correctly with respect to min_free_kbytes.
5899 void setup_per_zone_wmarks(void)
5901 mutex_lock(&zonelists_mutex);
5902 __setup_per_zone_wmarks();
5903 mutex_unlock(&zonelists_mutex);
5907 * The inactive anon list should be small enough that the VM never has to
5908 * do too much work, but large enough that each inactive page has a chance
5909 * to be referenced again before it is swapped out.
5911 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5912 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5913 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5914 * the anonymous pages are kept on the inactive list.
5917 * memory ratio inactive anon
5918 * -------------------------------------
5927 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5929 unsigned int gb, ratio;
5931 /* Zone size in gigabytes */
5932 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5934 ratio = int_sqrt(10 * gb);
5938 zone->inactive_ratio = ratio;
5941 static void __meminit setup_per_zone_inactive_ratio(void)
5946 calculate_zone_inactive_ratio(zone);
5950 * Initialise min_free_kbytes.
5952 * For small machines we want it small (128k min). For large machines
5953 * we want it large (64MB max). But it is not linear, because network
5954 * bandwidth does not increase linearly with machine size. We use
5956 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5957 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5973 int __meminit init_per_zone_wmark_min(void)
5975 unsigned long lowmem_kbytes;
5976 int new_min_free_kbytes;
5978 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5979 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5981 if (new_min_free_kbytes > user_min_free_kbytes) {
5982 min_free_kbytes = new_min_free_kbytes;
5983 if (min_free_kbytes < 128)
5984 min_free_kbytes = 128;
5985 if (min_free_kbytes > 65536)
5986 min_free_kbytes = 65536;
5988 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5989 new_min_free_kbytes, user_min_free_kbytes);
5991 setup_per_zone_wmarks();
5992 refresh_zone_stat_thresholds();
5993 setup_per_zone_lowmem_reserve();
5994 setup_per_zone_inactive_ratio();
5997 module_init(init_per_zone_wmark_min)
6000 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6001 * that we can call two helper functions whenever min_free_kbytes
6004 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6005 void __user *buffer, size_t *length, loff_t *ppos)
6009 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6014 user_min_free_kbytes = min_free_kbytes;
6015 setup_per_zone_wmarks();
6021 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6022 void __user *buffer, size_t *length, loff_t *ppos)
6027 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6032 zone->min_unmapped_pages = (zone->managed_pages *
6033 sysctl_min_unmapped_ratio) / 100;
6037 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6038 void __user *buffer, size_t *length, loff_t *ppos)
6043 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6048 zone->min_slab_pages = (zone->managed_pages *
6049 sysctl_min_slab_ratio) / 100;
6055 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6056 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6057 * whenever sysctl_lowmem_reserve_ratio changes.
6059 * The reserve ratio obviously has absolutely no relation with the
6060 * minimum watermarks. The lowmem reserve ratio can only make sense
6061 * if in function of the boot time zone sizes.
6063 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6064 void __user *buffer, size_t *length, loff_t *ppos)
6066 proc_dointvec_minmax(table, write, buffer, length, ppos);
6067 setup_per_zone_lowmem_reserve();
6072 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6073 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6074 * pagelist can have before it gets flushed back to buddy allocator.
6076 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6077 void __user *buffer, size_t *length, loff_t *ppos)
6080 int old_percpu_pagelist_fraction;
6083 mutex_lock(&pcp_batch_high_lock);
6084 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6086 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6087 if (!write || ret < 0)
6090 /* Sanity checking to avoid pcp imbalance */
6091 if (percpu_pagelist_fraction &&
6092 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6093 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6099 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6102 for_each_populated_zone(zone) {
6105 for_each_possible_cpu(cpu)
6106 pageset_set_high_and_batch(zone,
6107 per_cpu_ptr(zone->pageset, cpu));
6110 mutex_unlock(&pcp_batch_high_lock);
6114 int hashdist = HASHDIST_DEFAULT;
6117 static int __init set_hashdist(char *str)
6121 hashdist = simple_strtoul(str, &str, 0);
6124 __setup("hashdist=", set_hashdist);
6128 * allocate a large system hash table from bootmem
6129 * - it is assumed that the hash table must contain an exact power-of-2
6130 * quantity of entries
6131 * - limit is the number of hash buckets, not the total allocation size
6133 void *__init alloc_large_system_hash(const char *tablename,
6134 unsigned long bucketsize,
6135 unsigned long numentries,
6138 unsigned int *_hash_shift,
6139 unsigned int *_hash_mask,
6140 unsigned long low_limit,
6141 unsigned long high_limit)
6143 unsigned long long max = high_limit;
6144 unsigned long log2qty, size;
6147 /* allow the kernel cmdline to have a say */
6149 /* round applicable memory size up to nearest megabyte */
6150 numentries = nr_kernel_pages;
6152 /* It isn't necessary when PAGE_SIZE >= 1MB */
6153 if (PAGE_SHIFT < 20)
6154 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6156 /* limit to 1 bucket per 2^scale bytes of low memory */
6157 if (scale > PAGE_SHIFT)
6158 numentries >>= (scale - PAGE_SHIFT);
6160 numentries <<= (PAGE_SHIFT - scale);
6162 /* Make sure we've got at least a 0-order allocation.. */
6163 if (unlikely(flags & HASH_SMALL)) {
6164 /* Makes no sense without HASH_EARLY */
6165 WARN_ON(!(flags & HASH_EARLY));
6166 if (!(numentries >> *_hash_shift)) {
6167 numentries = 1UL << *_hash_shift;
6168 BUG_ON(!numentries);
6170 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6171 numentries = PAGE_SIZE / bucketsize;
6173 numentries = roundup_pow_of_two(numentries);
6175 /* limit allocation size to 1/16 total memory by default */
6177 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6178 do_div(max, bucketsize);
6180 max = min(max, 0x80000000ULL);
6182 if (numentries < low_limit)
6183 numentries = low_limit;
6184 if (numentries > max)
6187 log2qty = ilog2(numentries);
6190 size = bucketsize << log2qty;
6191 if (flags & HASH_EARLY)
6192 table = memblock_virt_alloc_nopanic(size, 0);
6194 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6197 * If bucketsize is not a power-of-two, we may free
6198 * some pages at the end of hash table which
6199 * alloc_pages_exact() automatically does
6201 if (get_order(size) < MAX_ORDER) {
6202 table = alloc_pages_exact(size, GFP_ATOMIC);
6203 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6206 } while (!table && size > PAGE_SIZE && --log2qty);
6209 panic("Failed to allocate %s hash table\n", tablename);
6211 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6214 ilog2(size) - PAGE_SHIFT,
6218 *_hash_shift = log2qty;
6220 *_hash_mask = (1 << log2qty) - 1;
6225 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6226 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6229 #ifdef CONFIG_SPARSEMEM
6230 return __pfn_to_section(pfn)->pageblock_flags;
6232 return zone->pageblock_flags;
6233 #endif /* CONFIG_SPARSEMEM */
6236 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6238 #ifdef CONFIG_SPARSEMEM
6239 pfn &= (PAGES_PER_SECTION-1);
6240 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6242 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6243 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6244 #endif /* CONFIG_SPARSEMEM */
6248 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6249 * @page: The page within the block of interest
6250 * @pfn: The target page frame number
6251 * @end_bitidx: The last bit of interest to retrieve
6252 * @mask: mask of bits that the caller is interested in
6254 * Return: pageblock_bits flags
6256 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6257 unsigned long end_bitidx,
6261 unsigned long *bitmap;
6262 unsigned long bitidx, word_bitidx;
6265 zone = page_zone(page);
6266 bitmap = get_pageblock_bitmap(zone, pfn);
6267 bitidx = pfn_to_bitidx(zone, pfn);
6268 word_bitidx = bitidx / BITS_PER_LONG;
6269 bitidx &= (BITS_PER_LONG-1);
6271 word = bitmap[word_bitidx];
6272 bitidx += end_bitidx;
6273 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6277 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6278 * @page: The page within the block of interest
6279 * @flags: The flags to set
6280 * @pfn: The target page frame number
6281 * @end_bitidx: The last bit of interest
6282 * @mask: mask of bits that the caller is interested in
6284 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6286 unsigned long end_bitidx,
6290 unsigned long *bitmap;
6291 unsigned long bitidx, word_bitidx;
6292 unsigned long old_word, word;
6294 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6296 zone = page_zone(page);
6297 bitmap = get_pageblock_bitmap(zone, pfn);
6298 bitidx = pfn_to_bitidx(zone, pfn);
6299 word_bitidx = bitidx / BITS_PER_LONG;
6300 bitidx &= (BITS_PER_LONG-1);
6302 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6304 bitidx += end_bitidx;
6305 mask <<= (BITS_PER_LONG - bitidx - 1);
6306 flags <<= (BITS_PER_LONG - bitidx - 1);
6308 word = READ_ONCE(bitmap[word_bitidx]);
6310 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6311 if (word == old_word)
6318 * This function checks whether pageblock includes unmovable pages or not.
6319 * If @count is not zero, it is okay to include less @count unmovable pages
6321 * PageLRU check without isolation or lru_lock could race so that
6322 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6323 * expect this function should be exact.
6325 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6326 bool skip_hwpoisoned_pages)
6328 unsigned long pfn, iter, found;
6332 * For avoiding noise data, lru_add_drain_all() should be called
6333 * If ZONE_MOVABLE, the zone never contains unmovable pages
6335 if (zone_idx(zone) == ZONE_MOVABLE)
6337 mt = get_pageblock_migratetype(page);
6338 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6341 pfn = page_to_pfn(page);
6342 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6343 unsigned long check = pfn + iter;
6345 if (!pfn_valid_within(check))
6348 page = pfn_to_page(check);
6351 * Hugepages are not in LRU lists, but they're movable.
6352 * We need not scan over tail pages bacause we don't
6353 * handle each tail page individually in migration.
6355 if (PageHuge(page)) {
6356 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6361 * We can't use page_count without pin a page
6362 * because another CPU can free compound page.
6363 * This check already skips compound tails of THP
6364 * because their page->_count is zero at all time.
6366 if (!atomic_read(&page->_count)) {
6367 if (PageBuddy(page))
6368 iter += (1 << page_order(page)) - 1;
6373 * The HWPoisoned page may be not in buddy system, and
6374 * page_count() is not 0.
6376 if (skip_hwpoisoned_pages && PageHWPoison(page))
6382 * If there are RECLAIMABLE pages, we need to check
6383 * it. But now, memory offline itself doesn't call
6384 * shrink_node_slabs() and it still to be fixed.
6387 * If the page is not RAM, page_count()should be 0.
6388 * we don't need more check. This is an _used_ not-movable page.
6390 * The problematic thing here is PG_reserved pages. PG_reserved
6391 * is set to both of a memory hole page and a _used_ kernel
6400 bool is_pageblock_removable_nolock(struct page *page)
6406 * We have to be careful here because we are iterating over memory
6407 * sections which are not zone aware so we might end up outside of
6408 * the zone but still within the section.
6409 * We have to take care about the node as well. If the node is offline
6410 * its NODE_DATA will be NULL - see page_zone.
6412 if (!node_online(page_to_nid(page)))
6415 zone = page_zone(page);
6416 pfn = page_to_pfn(page);
6417 if (!zone_spans_pfn(zone, pfn))
6420 return !has_unmovable_pages(zone, page, 0, true);
6425 static unsigned long pfn_max_align_down(unsigned long pfn)
6427 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6428 pageblock_nr_pages) - 1);
6431 static unsigned long pfn_max_align_up(unsigned long pfn)
6433 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6434 pageblock_nr_pages));
6437 /* [start, end) must belong to a single zone. */
6438 static int __alloc_contig_migrate_range(struct compact_control *cc,
6439 unsigned long start, unsigned long end)
6441 /* This function is based on compact_zone() from compaction.c. */
6442 unsigned long nr_reclaimed;
6443 unsigned long pfn = start;
6444 unsigned int tries = 0;
6449 while (pfn < end || !list_empty(&cc->migratepages)) {
6450 if (fatal_signal_pending(current)) {
6455 if (list_empty(&cc->migratepages)) {
6456 cc->nr_migratepages = 0;
6457 pfn = isolate_migratepages_range(cc, pfn, end);
6463 } else if (++tries == 5) {
6464 ret = ret < 0 ? ret : -EBUSY;
6468 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6470 cc->nr_migratepages -= nr_reclaimed;
6472 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6473 NULL, 0, cc->mode, MR_CMA);
6476 putback_movable_pages(&cc->migratepages);
6483 * alloc_contig_range() -- tries to allocate given range of pages
6484 * @start: start PFN to allocate
6485 * @end: one-past-the-last PFN to allocate
6486 * @migratetype: migratetype of the underlaying pageblocks (either
6487 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6488 * in range must have the same migratetype and it must
6489 * be either of the two.
6491 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6492 * aligned, however it's the caller's responsibility to guarantee that
6493 * we are the only thread that changes migrate type of pageblocks the
6496 * The PFN range must belong to a single zone.
6498 * Returns zero on success or negative error code. On success all
6499 * pages which PFN is in [start, end) are allocated for the caller and
6500 * need to be freed with free_contig_range().
6502 int alloc_contig_range(unsigned long start, unsigned long end,
6503 unsigned migratetype)
6505 unsigned long outer_start, outer_end;
6508 struct compact_control cc = {
6509 .nr_migratepages = 0,
6511 .zone = page_zone(pfn_to_page(start)),
6512 .mode = MIGRATE_SYNC,
6513 .ignore_skip_hint = true,
6515 INIT_LIST_HEAD(&cc.migratepages);
6518 * What we do here is we mark all pageblocks in range as
6519 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6520 * have different sizes, and due to the way page allocator
6521 * work, we align the range to biggest of the two pages so
6522 * that page allocator won't try to merge buddies from
6523 * different pageblocks and change MIGRATE_ISOLATE to some
6524 * other migration type.
6526 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6527 * migrate the pages from an unaligned range (ie. pages that
6528 * we are interested in). This will put all the pages in
6529 * range back to page allocator as MIGRATE_ISOLATE.
6531 * When this is done, we take the pages in range from page
6532 * allocator removing them from the buddy system. This way
6533 * page allocator will never consider using them.
6535 * This lets us mark the pageblocks back as
6536 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6537 * aligned range but not in the unaligned, original range are
6538 * put back to page allocator so that buddy can use them.
6541 ret = start_isolate_page_range(pfn_max_align_down(start),
6542 pfn_max_align_up(end), migratetype,
6547 ret = __alloc_contig_migrate_range(&cc, start, end);
6552 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6553 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6554 * more, all pages in [start, end) are free in page allocator.
6555 * What we are going to do is to allocate all pages from
6556 * [start, end) (that is remove them from page allocator).
6558 * The only problem is that pages at the beginning and at the
6559 * end of interesting range may be not aligned with pages that
6560 * page allocator holds, ie. they can be part of higher order
6561 * pages. Because of this, we reserve the bigger range and
6562 * once this is done free the pages we are not interested in.
6564 * We don't have to hold zone->lock here because the pages are
6565 * isolated thus they won't get removed from buddy.
6568 lru_add_drain_all();
6569 drain_all_pages(cc.zone);
6572 outer_start = start;
6573 while (!PageBuddy(pfn_to_page(outer_start))) {
6574 if (++order >= MAX_ORDER) {
6578 outer_start &= ~0UL << order;
6581 /* Make sure the range is really isolated. */
6582 if (test_pages_isolated(outer_start, end, false)) {
6583 pr_info("%s: [%lx, %lx) PFNs busy\n",
6584 __func__, outer_start, end);
6589 /* Grab isolated pages from freelists. */
6590 outer_end = isolate_freepages_range(&cc, outer_start, end);
6596 /* Free head and tail (if any) */
6597 if (start != outer_start)
6598 free_contig_range(outer_start, start - outer_start);
6599 if (end != outer_end)
6600 free_contig_range(end, outer_end - end);
6603 undo_isolate_page_range(pfn_max_align_down(start),
6604 pfn_max_align_up(end), migratetype);
6608 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6610 unsigned int count = 0;
6612 for (; nr_pages--; pfn++) {
6613 struct page *page = pfn_to_page(pfn);
6615 count += page_count(page) != 1;
6618 WARN(count != 0, "%d pages are still in use!\n", count);
6622 #ifdef CONFIG_MEMORY_HOTPLUG
6624 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6625 * page high values need to be recalulated.
6627 void __meminit zone_pcp_update(struct zone *zone)
6630 mutex_lock(&pcp_batch_high_lock);
6631 for_each_possible_cpu(cpu)
6632 pageset_set_high_and_batch(zone,
6633 per_cpu_ptr(zone->pageset, cpu));
6634 mutex_unlock(&pcp_batch_high_lock);
6638 void zone_pcp_reset(struct zone *zone)
6640 unsigned long flags;
6642 struct per_cpu_pageset *pset;
6644 /* avoid races with drain_pages() */
6645 local_irq_save(flags);
6646 if (zone->pageset != &boot_pageset) {
6647 for_each_online_cpu(cpu) {
6648 pset = per_cpu_ptr(zone->pageset, cpu);
6649 drain_zonestat(zone, pset);
6651 free_percpu(zone->pageset);
6652 zone->pageset = &boot_pageset;
6654 local_irq_restore(flags);
6657 #ifdef CONFIG_MEMORY_HOTREMOVE
6659 * All pages in the range must be isolated before calling this.
6662 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6666 unsigned int order, i;
6668 unsigned long flags;
6669 /* find the first valid pfn */
6670 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6675 zone = page_zone(pfn_to_page(pfn));
6676 spin_lock_irqsave(&zone->lock, flags);
6678 while (pfn < end_pfn) {
6679 if (!pfn_valid(pfn)) {
6683 page = pfn_to_page(pfn);
6685 * The HWPoisoned page may be not in buddy system, and
6686 * page_count() is not 0.
6688 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6690 SetPageReserved(page);
6694 BUG_ON(page_count(page));
6695 BUG_ON(!PageBuddy(page));
6696 order = page_order(page);
6697 #ifdef CONFIG_DEBUG_VM
6698 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6699 pfn, 1 << order, end_pfn);
6701 list_del(&page->lru);
6702 rmv_page_order(page);
6703 zone->free_area[order].nr_free--;
6704 for (i = 0; i < (1 << order); i++)
6705 SetPageReserved((page+i));
6706 pfn += (1 << order);
6708 spin_unlock_irqrestore(&zone->lock, flags);
6712 #ifdef CONFIG_MEMORY_FAILURE
6713 bool is_free_buddy_page(struct page *page)
6715 struct zone *zone = page_zone(page);
6716 unsigned long pfn = page_to_pfn(page);
6717 unsigned long flags;
6720 spin_lock_irqsave(&zone->lock, flags);
6721 for (order = 0; order < MAX_ORDER; order++) {
6722 struct page *page_head = page - (pfn & ((1 << order) - 1));
6724 if (PageBuddy(page_head) && page_order(page_head) >= order)
6727 spin_unlock_irqrestore(&zone->lock, flags);
6729 return order < MAX_ORDER;